Actual source code: mpibaij.c
1: /*$Id: mpibaij.c,v 1.234 2001/09/25 22:56:49 balay Exp $*/
3: #include src/mat/impls/baij/mpi/mpibaij.h
4: #include src/vec/vecimpl.h
6: EXTERN int MatSetUpMultiply_MPIBAIJ(Mat);
7: EXTERN int DisAssemble_MPIBAIJ(Mat);
8: EXTERN int MatIncreaseOverlap_MPIBAIJ(Mat,int,IS[],int);
9: EXTERN int MatGetSubMatrices_MPIBAIJ(Mat,int,const IS[],const IS[],MatReuse,Mat *[]);
10: EXTERN int MatGetValues_SeqBAIJ(Mat,int,const int[],int,const int [],PetscScalar []);
11: EXTERN int MatSetValues_SeqBAIJ(Mat,int,const int[],int,const int [],const PetscScalar [],InsertMode);
12: EXTERN int MatSetValuesBlocked_SeqBAIJ(Mat,int,const int[],int,const int[],const PetscScalar[],InsertMode);
13: EXTERN int MatGetRow_SeqBAIJ(Mat,int,int*,int*[],PetscScalar*[]);
14: EXTERN int MatRestoreRow_SeqBAIJ(Mat,int,int*,int*[],PetscScalar*[]);
15: EXTERN int MatPrintHelp_SeqBAIJ(Mat);
16: EXTERN int MatZeroRows_SeqBAIJ(Mat,IS,const PetscScalar*);
18: /* UGLY, ugly, ugly
19: When MatScalar == PetscScalar the function MatSetValuesBlocked_MPIBAIJ_MatScalar() does
20: not exist. Otherwise ..._MatScalar() takes matrix elements in single precision and
21: inserts them into the single precision data structure. The function MatSetValuesBlocked_MPIBAIJ()
22: converts the entries into single precision and then calls ..._MatScalar() to put them
23: into the single precision data structures.
24: */
25: #if defined(PETSC_USE_MAT_SINGLE)
26: EXTERN int MatSetValuesBlocked_SeqBAIJ_MatScalar(Mat,int,const int*,int,const int*,const MatScalar*,InsertMode);
27: EXTERN int MatSetValues_MPIBAIJ_MatScalar(Mat,int,const int*,int,const int*,const MatScalar*,InsertMode);
28: EXTERN int MatSetValuesBlocked_MPIBAIJ_MatScalar(Mat,int,const int*,int,const int*,const MatScalar*,InsertMode);
29: EXTERN int MatSetValues_MPIBAIJ_HT_MatScalar(Mat,int,const int*,int,const int*,const MatScalar*,InsertMode);
30: EXTERN int MatSetValuesBlocked_MPIBAIJ_HT_MatScalar(Mat,int,const int*,int,const int*,const MatScalar*,InsertMode);
31: #else
32: #define MatSetValuesBlocked_SeqBAIJ_MatScalar MatSetValuesBlocked_SeqBAIJ
33: #define MatSetValues_MPIBAIJ_MatScalar MatSetValues_MPIBAIJ
34: #define MatSetValuesBlocked_MPIBAIJ_MatScalar MatSetValuesBlocked_MPIBAIJ
35: #define MatSetValues_MPIBAIJ_HT_MatScalar MatSetValues_MPIBAIJ_HT
36: #define MatSetValuesBlocked_MPIBAIJ_HT_MatScalar MatSetValuesBlocked_MPIBAIJ_HT
37: #endif
41: int MatGetRowMax_MPIBAIJ(Mat A,Vec v)
42: {
43: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
44: int ierr,i;
45: PetscScalar *va,*vb;
46: Vec vtmp;
49:
50: MatGetRowMax(a->A,v);
51: VecGetArray(v,&va);
53: VecCreateSeq(PETSC_COMM_SELF,A->m,&vtmp);
54: MatGetRowMax(a->B,vtmp);
55: VecGetArray(vtmp,&vb);
57: for (i=0; i<A->m; i++){
58: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) va[i] = vb[i];
59: }
61: VecRestoreArray(v,&va);
62: VecRestoreArray(vtmp,&vb);
63: VecDestroy(vtmp);
64:
65: return(0);
66: }
68: EXTERN_C_BEGIN
71: int MatStoreValues_MPIBAIJ(Mat mat)
72: {
73: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data;
74: int ierr;
77: MatStoreValues(aij->A);
78: MatStoreValues(aij->B);
79: return(0);
80: }
81: EXTERN_C_END
83: EXTERN_C_BEGIN
86: int MatRetrieveValues_MPIBAIJ(Mat mat)
87: {
88: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data;
89: int ierr;
92: MatRetrieveValues(aij->A);
93: MatRetrieveValues(aij->B);
94: return(0);
95: }
96: EXTERN_C_END
98: /*
99: Local utility routine that creates a mapping from the global column
100: number to the local number in the off-diagonal part of the local
101: storage of the matrix. This is done in a non scable way since the
102: length of colmap equals the global matrix length.
103: */
106: static int CreateColmap_MPIBAIJ_Private(Mat mat)
107: {
108: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
109: Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data;
110: int nbs = B->nbs,i,bs=B->bs,ierr;
113: #if defined (PETSC_USE_CTABLE)
114: PetscTableCreate(baij->nbs,&baij->colmap);
115: for (i=0; i<nbs; i++){
116: PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1);
117: }
118: #else
119: PetscMalloc((baij->Nbs+1)*sizeof(int),&baij->colmap);
120: PetscLogObjectMemory(mat,baij->Nbs*sizeof(int));
121: PetscMemzero(baij->colmap,baij->Nbs*sizeof(int));
122: for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1;
123: #endif
124: return(0);
125: }
127: #define CHUNKSIZE 10
129: #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \
130: { \
131: \
132: brow = row/bs; \
133: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
134: rmax = aimax[brow]; nrow = ailen[brow]; \
135: bcol = col/bs; \
136: ridx = row % bs; cidx = col % bs; \
137: low = 0; high = nrow; \
138: while (high-low > 3) { \
139: t = (low+high)/2; \
140: if (rp[t] > bcol) high = t; \
141: else low = t; \
142: } \
143: for (_i=low; _i<high; _i++) { \
144: if (rp[_i] > bcol) break; \
145: if (rp[_i] == bcol) { \
146: bap = ap + bs2*_i + bs*cidx + ridx; \
147: if (addv == ADD_VALUES) *bap += value; \
148: else *bap = value; \
149: goto a_noinsert; \
150: } \
151: } \
152: if (a->nonew == 1) goto a_noinsert; \
153: else if (a->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) into matrix", row, col); \
154: if (nrow >= rmax) { \
155: /* there is no extra room in row, therefore enlarge */ \
156: int new_nz = ai[a->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
157: MatScalar *new_a; \
158: \
159: if (a->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) in the matrix", row, col); \
160: \
161: /* malloc new storage space */ \
162: len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(a->mbs+1)*sizeof(int); \
163: PetscMalloc(len,&new_a); \
164: new_j = (int*)(new_a + bs2*new_nz); \
165: new_i = new_j + new_nz; \
166: \
167: /* copy over old data into new slots */ \
168: for (ii=0; ii<brow+1; ii++) {new_i[ii] = ai[ii];} \
169: for (ii=brow+1; ii<a->mbs+1; ii++) {new_i[ii] = ai[ii]+CHUNKSIZE;} \
170: PetscMemcpy(new_j,aj,(ai[brow]+nrow)*sizeof(int)); \
171: len = (new_nz - CHUNKSIZE - ai[brow] - nrow); \
172: PetscMemcpy(new_j+ai[brow]+nrow+CHUNKSIZE,aj+ai[brow]+nrow,len*sizeof(int)); \
173: PetscMemcpy(new_a,aa,(ai[brow]+nrow)*bs2*sizeof(MatScalar)); \
174: PetscMemzero(new_a+bs2*(ai[brow]+nrow),bs2*CHUNKSIZE*sizeof(PetscScalar)); \
175: PetscMemcpy(new_a+bs2*(ai[brow]+nrow+CHUNKSIZE), \
176: aa+bs2*(ai[brow]+nrow),bs2*len*sizeof(MatScalar)); \
177: /* free up old matrix storage */ \
178: PetscFree(a->a); \
179: if (!a->singlemalloc) { \
180: PetscFree(a->i); \
181: PetscFree(a->j);\
182: } \
183: aa = a->a = new_a; ai = a->i = new_i; aj = a->j = new_j; \
184: a->singlemalloc = PETSC_TRUE; \
185: \
186: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
187: rmax = aimax[brow] = aimax[brow] + CHUNKSIZE; \
188: PetscLogObjectMemory(A,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \
189: a->maxnz += bs2*CHUNKSIZE; \
190: a->reallocs++; \
191: a->nz++; \
192: } \
193: N = nrow++ - 1; \
194: /* shift up all the later entries in this row */ \
195: for (ii=N; ii>=_i; ii--) { \
196: rp[ii+1] = rp[ii]; \
197: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
198: } \
199: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); } \
200: rp[_i] = bcol; \
201: ap[bs2*_i + bs*cidx + ridx] = value; \
202: a_noinsert:; \
203: ailen[brow] = nrow; \
204: }
206: #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \
207: { \
208: brow = row/bs; \
209: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
210: rmax = bimax[brow]; nrow = bilen[brow]; \
211: bcol = col/bs; \
212: ridx = row % bs; cidx = col % bs; \
213: low = 0; high = nrow; \
214: while (high-low > 3) { \
215: t = (low+high)/2; \
216: if (rp[t] > bcol) high = t; \
217: else low = t; \
218: } \
219: for (_i=low; _i<high; _i++) { \
220: if (rp[_i] > bcol) break; \
221: if (rp[_i] == bcol) { \
222: bap = ap + bs2*_i + bs*cidx + ridx; \
223: if (addv == ADD_VALUES) *bap += value; \
224: else *bap = value; \
225: goto b_noinsert; \
226: } \
227: } \
228: if (b->nonew == 1) goto b_noinsert; \
229: else if (b->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) into matrix", row, col); \
230: if (nrow >= rmax) { \
231: /* there is no extra room in row, therefore enlarge */ \
232: int new_nz = bi[b->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
233: MatScalar *new_a; \
234: \
235: if (b->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) in the matrix", row, col); \
236: \
237: /* malloc new storage space */ \
238: len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(b->mbs+1)*sizeof(int); \
239: PetscMalloc(len,&new_a); \
240: new_j = (int*)(new_a + bs2*new_nz); \
241: new_i = new_j + new_nz; \
242: \
243: /* copy over old data into new slots */ \
244: for (ii=0; ii<brow+1; ii++) {new_i[ii] = bi[ii];} \
245: for (ii=brow+1; ii<b->mbs+1; ii++) {new_i[ii] = bi[ii]+CHUNKSIZE;} \
246: PetscMemcpy(new_j,bj,(bi[brow]+nrow)*sizeof(int)); \
247: len = (new_nz - CHUNKSIZE - bi[brow] - nrow); \
248: PetscMemcpy(new_j+bi[brow]+nrow+CHUNKSIZE,bj+bi[brow]+nrow,len*sizeof(int)); \
249: PetscMemcpy(new_a,ba,(bi[brow]+nrow)*bs2*sizeof(MatScalar)); \
250: PetscMemzero(new_a+bs2*(bi[brow]+nrow),bs2*CHUNKSIZE*sizeof(MatScalar)); \
251: PetscMemcpy(new_a+bs2*(bi[brow]+nrow+CHUNKSIZE), \
252: ba+bs2*(bi[brow]+nrow),bs2*len*sizeof(MatScalar)); \
253: /* free up old matrix storage */ \
254: PetscFree(b->a); \
255: if (!b->singlemalloc) { \
256: PetscFree(b->i); \
257: PetscFree(b->j); \
258: } \
259: ba = b->a = new_a; bi = b->i = new_i; bj = b->j = new_j; \
260: b->singlemalloc = PETSC_TRUE; \
261: \
262: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
263: rmax = bimax[brow] = bimax[brow] + CHUNKSIZE; \
264: PetscLogObjectMemory(B,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \
265: b->maxnz += bs2*CHUNKSIZE; \
266: b->reallocs++; \
267: b->nz++; \
268: } \
269: N = nrow++ - 1; \
270: /* shift up all the later entries in this row */ \
271: for (ii=N; ii>=_i; ii--) { \
272: rp[ii+1] = rp[ii]; \
273: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
274: } \
275: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));} \
276: rp[_i] = bcol; \
277: ap[bs2*_i + bs*cidx + ridx] = value; \
278: b_noinsert:; \
279: bilen[brow] = nrow; \
280: }
282: #if defined(PETSC_USE_MAT_SINGLE)
285: int MatSetValues_MPIBAIJ(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
286: {
287: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
288: int ierr,i,N = m*n;
289: MatScalar *vsingle;
292: if (N > b->setvalueslen) {
293: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
294: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
295: b->setvalueslen = N;
296: }
297: vsingle = b->setvaluescopy;
299: for (i=0; i<N; i++) {
300: vsingle[i] = v[i];
301: }
302: MatSetValues_MPIBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);
303: return(0);
304: }
308: int MatSetValuesBlocked_MPIBAIJ(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
309: {
310: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
311: int ierr,i,N = m*n*b->bs2;
312: MatScalar *vsingle;
315: if (N > b->setvalueslen) {
316: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
317: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
318: b->setvalueslen = N;
319: }
320: vsingle = b->setvaluescopy;
321: for (i=0; i<N; i++) {
322: vsingle[i] = v[i];
323: }
324: MatSetValuesBlocked_MPIBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);
325: return(0);
326: }
330: int MatSetValues_MPIBAIJ_HT(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
331: {
332: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
333: int ierr,i,N = m*n;
334: MatScalar *vsingle;
337: if (N > b->setvalueslen) {
338: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
339: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
340: b->setvalueslen = N;
341: }
342: vsingle = b->setvaluescopy;
343: for (i=0; i<N; i++) {
344: vsingle[i] = v[i];
345: }
346: MatSetValues_MPIBAIJ_HT_MatScalar(mat,m,im,n,in,vsingle,addv);
347: return(0);
348: }
352: int MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
353: {
354: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
355: int ierr,i,N = m*n*b->bs2;
356: MatScalar *vsingle;
359: if (N > b->setvalueslen) {
360: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
361: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
362: b->setvalueslen = N;
363: }
364: vsingle = b->setvaluescopy;
365: for (i=0; i<N; i++) {
366: vsingle[i] = v[i];
367: }
368: MatSetValuesBlocked_MPIBAIJ_HT_MatScalar(mat,m,im,n,in,vsingle,addv);
369: return(0);
370: }
371: #endif
375: int MatSetValues_MPIBAIJ_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
376: {
377: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
378: MatScalar value;
379: PetscTruth roworiented = baij->roworiented;
380: int ierr,i,j,row,col;
381: int rstart_orig=baij->rstart_bs;
382: int rend_orig=baij->rend_bs,cstart_orig=baij->cstart_bs;
383: int cend_orig=baij->cend_bs,bs=baij->bs;
385: /* Some Variables required in the macro */
386: Mat A = baij->A;
387: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)(A)->data;
388: int *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
389: MatScalar *aa=a->a;
391: Mat B = baij->B;
392: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data;
393: int *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
394: MatScalar *ba=b->a;
396: int *rp,ii,nrow,_i,rmax,N,brow,bcol;
397: int low,high,t,ridx,cidx,bs2=a->bs2;
398: MatScalar *ap,*bap;
401: for (i=0; i<m; i++) {
402: if (im[i] < 0) continue;
403: #if defined(PETSC_USE_BOPT_g)
404: if (im[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",im[i],mat->M-1);
405: #endif
406: if (im[i] >= rstart_orig && im[i] < rend_orig) {
407: row = im[i] - rstart_orig;
408: for (j=0; j<n; j++) {
409: if (in[j] >= cstart_orig && in[j] < cend_orig){
410: col = in[j] - cstart_orig;
411: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
412: MatSetValues_SeqBAIJ_A_Private(row,col,value,addv);
413: /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
414: } else if (in[j] < 0) continue;
415: #if defined(PETSC_USE_BOPT_g)
416: else if (in[j] >= mat->N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",in[i],mat->N-1);}
417: #endif
418: else {
419: if (mat->was_assembled) {
420: if (!baij->colmap) {
421: CreateColmap_MPIBAIJ_Private(mat);
422: }
423: #if defined (PETSC_USE_CTABLE)
424: PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
425: col = col - 1;
426: #else
427: col = baij->colmap[in[j]/bs] - 1;
428: #endif
429: if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
430: DisAssemble_MPIBAIJ(mat);
431: col = in[j];
432: /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
433: B = baij->B;
434: b = (Mat_SeqBAIJ*)(B)->data;
435: bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
436: ba=b->a;
437: } else col += in[j]%bs;
438: } else col = in[j];
439: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
440: MatSetValues_SeqBAIJ_B_Private(row,col,value,addv);
441: /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
442: }
443: }
444: } else {
445: if (!baij->donotstash) {
446: if (roworiented) {
447: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
448: } else {
449: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
450: }
451: }
452: }
453: }
454: return(0);
455: }
459: int MatSetValuesBlocked_MPIBAIJ_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
460: {
461: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
462: const MatScalar *value;
463: MatScalar *barray=baij->barray;
464: PetscTruth roworiented = baij->roworiented;
465: int ierr,i,j,ii,jj,row,col,rstart=baij->rstart;
466: int rend=baij->rend,cstart=baij->cstart,stepval;
467: int cend=baij->cend,bs=baij->bs,bs2=baij->bs2;
468:
470: if(!barray) {
471: PetscMalloc(bs2*sizeof(MatScalar),&barray);
472: baij->barray = barray;
473: }
475: if (roworiented) {
476: stepval = (n-1)*bs;
477: } else {
478: stepval = (m-1)*bs;
479: }
480: for (i=0; i<m; i++) {
481: if (im[i] < 0) continue;
482: #if defined(PETSC_USE_BOPT_g)
483: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %d max %d",im[i],baij->Mbs-1);
484: #endif
485: if (im[i] >= rstart && im[i] < rend) {
486: row = im[i] - rstart;
487: for (j=0; j<n; j++) {
488: /* If NumCol = 1 then a copy is not required */
489: if ((roworiented) && (n == 1)) {
490: barray = (MatScalar*)v + i*bs2;
491: } else if((!roworiented) && (m == 1)) {
492: barray = (MatScalar*)v + j*bs2;
493: } else { /* Here a copy is required */
494: if (roworiented) {
495: value = v + i*(stepval+bs)*bs + j*bs;
496: } else {
497: value = v + j*(stepval+bs)*bs + i*bs;
498: }
499: for (ii=0; ii<bs; ii++,value+=stepval) {
500: for (jj=0; jj<bs; jj++) {
501: *barray++ = *value++;
502: }
503: }
504: barray -=bs2;
505: }
506:
507: if (in[j] >= cstart && in[j] < cend){
508: col = in[j] - cstart;
509: MatSetValuesBlocked_SeqBAIJ_MatScalar(baij->A,1,&row,1,&col,barray,addv);
510: }
511: else if (in[j] < 0) continue;
512: #if defined(PETSC_USE_BOPT_g)
513: else if (in[j] >= baij->Nbs) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %d max %d",in[j],baij->Nbs-1);}
514: #endif
515: else {
516: if (mat->was_assembled) {
517: if (!baij->colmap) {
518: CreateColmap_MPIBAIJ_Private(mat);
519: }
521: #if defined(PETSC_USE_BOPT_g)
522: #if defined (PETSC_USE_CTABLE)
523: { int data;
524: PetscTableFind(baij->colmap,in[j]+1,&data);
525: if ((data - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
526: }
527: #else
528: if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
529: #endif
530: #endif
531: #if defined (PETSC_USE_CTABLE)
532: PetscTableFind(baij->colmap,in[j]+1,&col);
533: col = (col - 1)/bs;
534: #else
535: col = (baij->colmap[in[j]] - 1)/bs;
536: #endif
537: if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
538: DisAssemble_MPIBAIJ(mat);
539: col = in[j];
540: }
541: }
542: else col = in[j];
543: MatSetValuesBlocked_SeqBAIJ_MatScalar(baij->B,1,&row,1,&col,barray,addv);
544: }
545: }
546: } else {
547: if (!baij->donotstash) {
548: if (roworiented) {
549: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
550: } else {
551: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
552: }
553: }
554: }
555: }
556: return(0);
557: }
559: #define HASH_KEY 0.6180339887
560: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(int)((size)*(tmp-(int)tmp)))
561: /* #define HASH(size,key) ((int)((size)*fmod(((key)*HASH_KEY),1))) */
562: /* #define HASH(size,key,tmp) ((int)((size)*fmod(((key)*HASH_KEY),1))) */
565: int MatSetValues_MPIBAIJ_HT_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
566: {
567: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
568: PetscTruth roworiented = baij->roworiented;
569: int ierr,i,j,row,col;
570: int rstart_orig=baij->rstart_bs;
571: int rend_orig=baij->rend_bs,Nbs=baij->Nbs;
572: int h1,key,size=baij->ht_size,bs=baij->bs,*HT=baij->ht,idx;
573: PetscReal tmp;
574: MatScalar **HD = baij->hd,value;
575: #if defined(PETSC_USE_BOPT_g)
576: int total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
577: #endif
581: for (i=0; i<m; i++) {
582: #if defined(PETSC_USE_BOPT_g)
583: if (im[i] < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative row");
584: if (im[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",im[i],mat->M-1);
585: #endif
586: row = im[i];
587: if (row >= rstart_orig && row < rend_orig) {
588: for (j=0; j<n; j++) {
589: col = in[j];
590: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
591: /* Look up into the Hash Table */
592: key = (row/bs)*Nbs+(col/bs)+1;
593: h1 = HASH(size,key,tmp);
595:
596: idx = h1;
597: #if defined(PETSC_USE_BOPT_g)
598: insert_ct++;
599: total_ct++;
600: if (HT[idx] != key) {
601: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++);
602: if (idx == size) {
603: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++);
604: if (idx == h1) {
605: SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%d,%d) has no entry in the hash table", row, col);
606: }
607: }
608: }
609: #else
610: if (HT[idx] != key) {
611: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++);
612: if (idx == size) {
613: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++);
614: if (idx == h1) {
615: SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%d,%d) has no entry in the hash table", row, col);
616: }
617: }
618: }
619: #endif
620: /* A HASH table entry is found, so insert the values at the correct address */
621: if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value;
622: else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value;
623: }
624: } else {
625: if (!baij->donotstash) {
626: if (roworiented) {
627: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
628: } else {
629: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
630: }
631: }
632: }
633: }
634: #if defined(PETSC_USE_BOPT_g)
635: baij->ht_total_ct = total_ct;
636: baij->ht_insert_ct = insert_ct;
637: #endif
638: return(0);
639: }
643: int MatSetValuesBlocked_MPIBAIJ_HT_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
644: {
645: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
646: PetscTruth roworiented = baij->roworiented;
647: int ierr,i,j,ii,jj,row,col;
648: int rstart=baij->rstart ;
649: int rend=baij->rend,stepval,bs=baij->bs,bs2=baij->bs2;
650: int h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs;
651: PetscReal tmp;
652: MatScalar **HD = baij->hd,*baij_a;
653: const MatScalar *v_t,*value;
654: #if defined(PETSC_USE_BOPT_g)
655: int total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
656: #endif
657:
660: if (roworiented) {
661: stepval = (n-1)*bs;
662: } else {
663: stepval = (m-1)*bs;
664: }
665: for (i=0; i<m; i++) {
666: #if defined(PETSC_USE_BOPT_g)
667: if (im[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %d",im[i]);
668: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",im[i],baij->Mbs-1);
669: #endif
670: row = im[i];
671: v_t = v + i*bs2;
672: if (row >= rstart && row < rend) {
673: for (j=0; j<n; j++) {
674: col = in[j];
676: /* Look up into the Hash Table */
677: key = row*Nbs+col+1;
678: h1 = HASH(size,key,tmp);
679:
680: idx = h1;
681: #if defined(PETSC_USE_BOPT_g)
682: total_ct++;
683: insert_ct++;
684: if (HT[idx] != key) {
685: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++);
686: if (idx == size) {
687: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++);
688: if (idx == h1) {
689: SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%d,%d) has no entry in the hash table", row, col);
690: }
691: }
692: }
693: #else
694: if (HT[idx] != key) {
695: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++);
696: if (idx == size) {
697: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++);
698: if (idx == h1) {
699: SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"(%d,%d) has no entry in the hash table", row, col);
700: }
701: }
702: }
703: #endif
704: baij_a = HD[idx];
705: if (roworiented) {
706: /*value = v + i*(stepval+bs)*bs + j*bs;*/
707: /* value = v + (i*(stepval+bs)+j)*bs; */
708: value = v_t;
709: v_t += bs;
710: if (addv == ADD_VALUES) {
711: for (ii=0; ii<bs; ii++,value+=stepval) {
712: for (jj=ii; jj<bs2; jj+=bs) {
713: baij_a[jj] += *value++;
714: }
715: }
716: } else {
717: for (ii=0; ii<bs; ii++,value+=stepval) {
718: for (jj=ii; jj<bs2; jj+=bs) {
719: baij_a[jj] = *value++;
720: }
721: }
722: }
723: } else {
724: value = v + j*(stepval+bs)*bs + i*bs;
725: if (addv == ADD_VALUES) {
726: for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
727: for (jj=0; jj<bs; jj++) {
728: baij_a[jj] += *value++;
729: }
730: }
731: } else {
732: for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
733: for (jj=0; jj<bs; jj++) {
734: baij_a[jj] = *value++;
735: }
736: }
737: }
738: }
739: }
740: } else {
741: if (!baij->donotstash) {
742: if (roworiented) {
743: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
744: } else {
745: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
746: }
747: }
748: }
749: }
750: #if defined(PETSC_USE_BOPT_g)
751: baij->ht_total_ct = total_ct;
752: baij->ht_insert_ct = insert_ct;
753: #endif
754: return(0);
755: }
759: int MatGetValues_MPIBAIJ(Mat mat,int m,const int idxm[],int n,const int idxn[],PetscScalar v[])
760: {
761: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
762: int bs=baij->bs,ierr,i,j,bsrstart = baij->rstart*bs,bsrend = baij->rend*bs;
763: int bscstart = baij->cstart*bs,bscend = baij->cend*bs,row,col,data;
766: for (i=0; i<m; i++) {
767: if (idxm[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %d",idxm[i]);
768: if (idxm[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",idxm[i],mat->M-1);
769: if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
770: row = idxm[i] - bsrstart;
771: for (j=0; j<n; j++) {
772: if (idxn[j] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %d",idxn[j]);
773: if (idxn[j] >= mat->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",idxn[j],mat->N-1);
774: if (idxn[j] >= bscstart && idxn[j] < bscend){
775: col = idxn[j] - bscstart;
776: MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
777: } else {
778: if (!baij->colmap) {
779: CreateColmap_MPIBAIJ_Private(mat);
780: }
781: #if defined (PETSC_USE_CTABLE)
782: PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
783: data --;
784: #else
785: data = baij->colmap[idxn[j]/bs]-1;
786: #endif
787: if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
788: else {
789: col = data + idxn[j]%bs;
790: MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
791: }
792: }
793: }
794: } else {
795: SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
796: }
797: }
798: return(0);
799: }
803: int MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm)
804: {
805: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
806: Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data;
807: int ierr,i,bs2=baij->bs2;
808: PetscReal sum = 0.0;
809: MatScalar *v;
812: if (baij->size == 1) {
813: MatNorm(baij->A,type,nrm);
814: } else {
815: if (type == NORM_FROBENIUS) {
816: v = amat->a;
817: for (i=0; i<amat->nz*bs2; i++) {
818: #if defined(PETSC_USE_COMPLEX)
819: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
820: #else
821: sum += (*v)*(*v); v++;
822: #endif
823: }
824: v = bmat->a;
825: for (i=0; i<bmat->nz*bs2; i++) {
826: #if defined(PETSC_USE_COMPLEX)
827: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
828: #else
829: sum += (*v)*(*v); v++;
830: #endif
831: }
832: MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPI_SUM,mat->comm);
833: *nrm = sqrt(*nrm);
834: } else {
835: SETERRQ(PETSC_ERR_SUP,"No support for this norm yet");
836: }
837: }
838: return(0);
839: }
842: /*
843: Creates the hash table, and sets the table
844: This table is created only once.
845: If new entried need to be added to the matrix
846: then the hash table has to be destroyed and
847: recreated.
848: */
851: int MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor)
852: {
853: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
854: Mat A = baij->A,B=baij->B;
855: Mat_SeqBAIJ *a=(Mat_SeqBAIJ *)A->data,*b=(Mat_SeqBAIJ *)B->data;
856: int i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
857: int size,bs2=baij->bs2,rstart=baij->rstart,ierr;
858: int cstart=baij->cstart,*garray=baij->garray,row,col,Nbs=baij->Nbs;
859: int *HT,key;
860: MatScalar **HD;
861: PetscReal tmp;
862: #if defined(PETSC_USE_BOPT_g)
863: int ct=0,max=0;
864: #endif
867: baij->ht_size=(int)(factor*nz);
868: size = baij->ht_size;
870: if (baij->ht) {
871: return(0);
872: }
873:
874: /* Allocate Memory for Hash Table */
875: PetscMalloc((size)*(sizeof(int)+sizeof(MatScalar*))+1,&baij->hd);
876: baij->ht = (int*)(baij->hd + size);
877: HD = baij->hd;
878: HT = baij->ht;
881: PetscMemzero(HD,size*(sizeof(int)+sizeof(PetscScalar*)));
882:
884: /* Loop Over A */
885: for (i=0; i<a->mbs; i++) {
886: for (j=ai[i]; j<ai[i+1]; j++) {
887: row = i+rstart;
888: col = aj[j]+cstart;
889:
890: key = row*Nbs + col + 1;
891: h1 = HASH(size,key,tmp);
892: for (k=0; k<size; k++){
893: if (HT[(h1+k)%size] == 0.0) {
894: HT[(h1+k)%size] = key;
895: HD[(h1+k)%size] = a->a + j*bs2;
896: break;
897: #if defined(PETSC_USE_BOPT_g)
898: } else {
899: ct++;
900: #endif
901: }
902: }
903: #if defined(PETSC_USE_BOPT_g)
904: if (k> max) max = k;
905: #endif
906: }
907: }
908: /* Loop Over B */
909: for (i=0; i<b->mbs; i++) {
910: for (j=bi[i]; j<bi[i+1]; j++) {
911: row = i+rstart;
912: col = garray[bj[j]];
913: key = row*Nbs + col + 1;
914: h1 = HASH(size,key,tmp);
915: for (k=0; k<size; k++){
916: if (HT[(h1+k)%size] == 0.0) {
917: HT[(h1+k)%size] = key;
918: HD[(h1+k)%size] = b->a + j*bs2;
919: break;
920: #if defined(PETSC_USE_BOPT_g)
921: } else {
922: ct++;
923: #endif
924: }
925: }
926: #if defined(PETSC_USE_BOPT_g)
927: if (k> max) max = k;
928: #endif
929: }
930: }
931:
932: /* Print Summary */
933: #if defined(PETSC_USE_BOPT_g)
934: for (i=0,j=0; i<size; i++) {
935: if (HT[i]) {j++;}
936: }
937: PetscLogInfo(0,"MatCreateHashTable_MPIBAIJ_Private: Average Search = %5.2f,max search = %d\n",(j== 0)? 0.0:((PetscReal)(ct+j))/j,max);
938: #endif
939: return(0);
940: }
944: int MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode)
945: {
946: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
947: int ierr,nstash,reallocs;
948: InsertMode addv;
951: if (baij->donotstash) {
952: return(0);
953: }
955: /* make sure all processors are either in INSERTMODE or ADDMODE */
956: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);
957: if (addv == (ADD_VALUES|INSERT_VALUES)) {
958: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
959: }
960: mat->insertmode = addv; /* in case this processor had no cache */
962: MatStashScatterBegin_Private(&mat->stash,baij->rowners_bs);
963: MatStashScatterBegin_Private(&mat->bstash,baij->rowners);
964: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
965: PetscLogInfo(0,"MatAssemblyBegin_MPIBAIJ:Stash has %d entries,uses %d mallocs.\n",nstash,reallocs);
966: MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);
967: PetscLogInfo(0,"MatAssemblyBegin_MPIBAIJ:Block-Stash has %d entries, uses %d mallocs.\n",nstash,reallocs);
968: return(0);
969: }
973: int MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode)
974: {
975: Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data;
976: Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)baij->A->data,*b=(Mat_SeqBAIJ*)baij->B->data;
977: int i,j,rstart,ncols,n,ierr,flg,bs2=baij->bs2;
978: int *row,*col,other_disassembled;
979: PetscTruth r1,r2,r3;
980: MatScalar *val;
981: InsertMode addv = mat->insertmode;
984: if (!baij->donotstash) {
985: while (1) {
986: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
987: if (!flg) break;
989: for (i=0; i<n;) {
990: /* Now identify the consecutive vals belonging to the same row */
991: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
992: if (j < n) ncols = j-i;
993: else ncols = n-i;
994: /* Now assemble all these values with a single function call */
995: MatSetValues_MPIBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i,addv);
996: i = j;
997: }
998: }
999: MatStashScatterEnd_Private(&mat->stash);
1000: /* Now process the block-stash. Since the values are stashed column-oriented,
1001: set the roworiented flag to column oriented, and after MatSetValues()
1002: restore the original flags */
1003: r1 = baij->roworiented;
1004: r2 = a->roworiented;
1005: r3 = b->roworiented;
1006: baij->roworiented = PETSC_FALSE;
1007: a->roworiented = PETSC_FALSE;
1008: b->roworiented = PETSC_FALSE;
1009: while (1) {
1010: MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
1011: if (!flg) break;
1012:
1013: for (i=0; i<n;) {
1014: /* Now identify the consecutive vals belonging to the same row */
1015: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
1016: if (j < n) ncols = j-i;
1017: else ncols = n-i;
1018: MatSetValuesBlocked_MPIBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
1019: i = j;
1020: }
1021: }
1022: MatStashScatterEnd_Private(&mat->bstash);
1023: baij->roworiented = r1;
1024: a->roworiented = r2;
1025: b->roworiented = r3;
1026: }
1028: MatAssemblyBegin(baij->A,mode);
1029: MatAssemblyEnd(baij->A,mode);
1031: /* determine if any processor has disassembled, if so we must
1032: also disassemble ourselfs, in order that we may reassemble. */
1033: /*
1034: if nonzero structure of submatrix B cannot change then we know that
1035: no processor disassembled thus we can skip this stuff
1036: */
1037: if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
1038: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);
1039: if (mat->was_assembled && !other_disassembled) {
1040: DisAssemble_MPIBAIJ(mat);
1041: }
1042: }
1044: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
1045: MatSetUpMultiply_MPIBAIJ(mat);
1046: }
1047: MatAssemblyBegin(baij->B,mode);
1048: MatAssemblyEnd(baij->B,mode);
1049:
1050: #if defined(PETSC_USE_BOPT_g)
1051: if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
1052: PetscLogInfo(0,"MatAssemblyEnd_MPIBAIJ:Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);
1053: baij->ht_total_ct = 0;
1054: baij->ht_insert_ct = 0;
1055: }
1056: #endif
1057: if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
1058: MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);
1059: mat->ops->setvalues = MatSetValues_MPIBAIJ_HT;
1060: mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT;
1061: }
1063: if (baij->rowvalues) {
1064: PetscFree(baij->rowvalues);
1065: baij->rowvalues = 0;
1066: }
1067: return(0);
1068: }
1072: static int MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1073: {
1074: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1075: int ierr,bs = baij->bs,size = baij->size,rank = baij->rank;
1076: PetscTruth isascii,isdraw;
1077: PetscViewer sviewer;
1078: PetscViewerFormat format;
1081: /* printf(" MatView_MPIBAIJ_ASCIIorDraworSocket is called ...\n"); */
1082: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);
1083: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
1084: if (isascii) {
1085: PetscViewerGetFormat(viewer,&format);
1086: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1087: MatInfo info;
1088: MPI_Comm_rank(mat->comm,&rank);
1089: MatGetInfo(mat,MAT_LOCAL,&info);
1090: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %d nz %d nz alloced %d bs %d mem %d\n",
1091: rank,mat->m,(int)info.nz_used*bs,(int)info.nz_allocated*bs,
1092: baij->bs,(int)info.memory);
1093: MatGetInfo(baij->A,MAT_LOCAL,&info);
1094: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);
1095: MatGetInfo(baij->B,MAT_LOCAL,&info);
1096: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);
1097: PetscViewerFlush(viewer);
1098: VecScatterView(baij->Mvctx,viewer);
1099: return(0);
1100: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1101: PetscViewerASCIIPrintf(viewer," block size is %d\n",bs);
1102: return(0);
1103: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1104: return(0);
1105: }
1106: }
1108: if (isdraw) {
1109: PetscDraw draw;
1110: PetscTruth isnull;
1111: PetscViewerDrawGetDraw(viewer,0,&draw);
1112: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
1113: }
1115: if (size == 1) {
1116: PetscObjectSetName((PetscObject)baij->A,mat->name);
1117: MatView(baij->A,viewer);
1118: } else {
1119: /* assemble the entire matrix onto first processor. */
1120: Mat A;
1121: Mat_SeqBAIJ *Aloc;
1122: int M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
1123: MatScalar *a;
1125: if (!rank) {
1126: MatCreateMPIBAIJ(mat->comm,baij->bs,M,N,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);
1127: } else {
1128: MatCreateMPIBAIJ(mat->comm,baij->bs,0,0,M,N,0,PETSC_NULL,0,PETSC_NULL,&A);
1129: }
1130: PetscLogObjectParent(mat,A);
1132: /* copy over the A part */
1133: Aloc = (Mat_SeqBAIJ*)baij->A->data;
1134: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1135: PetscMalloc(bs*sizeof(int),&rvals);
1137: for (i=0; i<mbs; i++) {
1138: rvals[0] = bs*(baij->rstart + i);
1139: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
1140: for (j=ai[i]; j<ai[i+1]; j++) {
1141: col = (baij->cstart+aj[j])*bs;
1142: for (k=0; k<bs; k++) {
1143: MatSetValues_MPIBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
1144: col++; a += bs;
1145: }
1146: }
1147: }
1148: /* copy over the B part */
1149: Aloc = (Mat_SeqBAIJ*)baij->B->data;
1150: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1151: for (i=0; i<mbs; i++) {
1152: rvals[0] = bs*(baij->rstart + i);
1153: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
1154: for (j=ai[i]; j<ai[i+1]; j++) {
1155: col = baij->garray[aj[j]]*bs;
1156: for (k=0; k<bs; k++) {
1157: MatSetValues_MPIBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
1158: col++; a += bs;
1159: }
1160: }
1161: }
1162: PetscFree(rvals);
1163: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1164: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1165: /*
1166: Everyone has to call to draw the matrix since the graphics waits are
1167: synchronized across all processors that share the PetscDraw object
1168: */
1169: PetscViewerGetSingleton(viewer,&sviewer);
1170: if (!rank) {
1171: PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,mat->name);
1172: MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);
1173: }
1174: PetscViewerRestoreSingleton(viewer,&sviewer);
1175: MatDestroy(A);
1176: }
1177: return(0);
1178: }
1182: int MatView_MPIBAIJ(Mat mat,PetscViewer viewer)
1183: {
1184: int ierr;
1185: PetscTruth isascii,isdraw,issocket,isbinary;
1188: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);
1189: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
1190: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
1191: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
1192: if (isascii || isdraw || issocket || isbinary) {
1193: MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);
1194: } else {
1195: SETERRQ1(1,"Viewer type %s not supported by MPIBAIJ matrices",((PetscObject)viewer)->type_name);
1196: }
1197: return(0);
1198: }
1202: int MatDestroy_MPIBAIJ(Mat mat)
1203: {
1204: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1205: int ierr;
1208: #if defined(PETSC_USE_LOG)
1209: PetscLogObjectState((PetscObject)mat,"Rows=%d,Cols=%d",mat->M,mat->N);
1210: #endif
1211: MatStashDestroy_Private(&mat->stash);
1212: MatStashDestroy_Private(&mat->bstash);
1213: PetscFree(baij->rowners);
1214: MatDestroy(baij->A);
1215: MatDestroy(baij->B);
1216: #if defined (PETSC_USE_CTABLE)
1217: if (baij->colmap) {PetscTableDelete(baij->colmap);}
1218: #else
1219: if (baij->colmap) {PetscFree(baij->colmap);}
1220: #endif
1221: if (baij->garray) {PetscFree(baij->garray);}
1222: if (baij->lvec) {VecDestroy(baij->lvec);}
1223: if (baij->Mvctx) {VecScatterDestroy(baij->Mvctx);}
1224: if (baij->rowvalues) {PetscFree(baij->rowvalues);}
1225: if (baij->barray) {PetscFree(baij->barray);}
1226: if (baij->hd) {PetscFree(baij->hd);}
1227: #if defined(PETSC_USE_MAT_SINGLE)
1228: if (baij->setvaluescopy) {PetscFree(baij->setvaluescopy);}
1229: #endif
1230: PetscFree(baij);
1231: return(0);
1232: }
1236: int MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy)
1237: {
1238: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1239: int ierr,nt;
1242: VecGetLocalSize(xx,&nt);
1243: if (nt != A->n) {
1244: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1245: }
1246: VecGetLocalSize(yy,&nt);
1247: if (nt != A->m) {
1248: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1249: }
1250: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1251: (*a->A->ops->mult)(a->A,xx,yy);
1252: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1253: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1254: VecScatterPostRecvs(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1255: return(0);
1256: }
1260: int MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1261: {
1262: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1263: int ierr;
1266: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1267: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1268: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1269: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1270: return(0);
1271: }
1275: int MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy)
1276: {
1277: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1278: int ierr;
1281: /* do nondiagonal part */
1282: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1283: /* send it on its way */
1284: VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1285: /* do local part */
1286: (*a->A->ops->multtranspose)(a->A,xx,yy);
1287: /* receive remote parts: note this assumes the values are not actually */
1288: /* inserted in yy until the next line, which is true for my implementation*/
1289: /* but is not perhaps always true. */
1290: VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1291: return(0);
1292: }
1296: int MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1297: {
1298: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1299: int ierr;
1302: /* do nondiagonal part */
1303: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1304: /* send it on its way */
1305: VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1306: /* do local part */
1307: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1308: /* receive remote parts: note this assumes the values are not actually */
1309: /* inserted in yy until the next line, which is true for my implementation*/
1310: /* but is not perhaps always true. */
1311: VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1312: return(0);
1313: }
1315: /*
1316: This only works correctly for square matrices where the subblock A->A is the
1317: diagonal block
1318: */
1321: int MatGetDiagonal_MPIBAIJ(Mat A,Vec v)
1322: {
1323: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1324: int ierr;
1327: if (A->M != A->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1328: MatGetDiagonal(a->A,v);
1329: return(0);
1330: }
1334: int MatScale_MPIBAIJ(const PetscScalar *aa,Mat A)
1335: {
1336: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1337: int ierr;
1340: MatScale(aa,a->A);
1341: MatScale(aa,a->B);
1342: return(0);
1343: }
1347: int MatGetRow_MPIBAIJ(Mat matin,int row,int *nz,int **idx,PetscScalar **v)
1348: {
1349: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data;
1350: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1351: int bs = mat->bs,bs2 = mat->bs2,i,ierr,*cworkA,*cworkB,**pcA,**pcB;
1352: int nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs;
1353: int *cmap,*idx_p,cstart = mat->cstart;
1356: if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1357: mat->getrowactive = PETSC_TRUE;
1359: if (!mat->rowvalues && (idx || v)) {
1360: /*
1361: allocate enough space to hold information from the longest row.
1362: */
1363: Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data;
1364: int max = 1,mbs = mat->mbs,tmp;
1365: for (i=0; i<mbs; i++) {
1366: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1367: if (max < tmp) { max = tmp; }
1368: }
1369: PetscMalloc(max*bs2*(sizeof(int)+sizeof(PetscScalar)),&mat->rowvalues);
1370: mat->rowindices = (int*)(mat->rowvalues + max*bs2);
1371: }
1372:
1373: if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows")
1374: lrow = row - brstart;
1376: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1377: if (!v) {pvA = 0; pvB = 0;}
1378: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1379: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1380: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1381: nztot = nzA + nzB;
1383: cmap = mat->garray;
1384: if (v || idx) {
1385: if (nztot) {
1386: /* Sort by increasing column numbers, assuming A and B already sorted */
1387: int imark = -1;
1388: if (v) {
1389: *v = v_p = mat->rowvalues;
1390: for (i=0; i<nzB; i++) {
1391: if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1392: else break;
1393: }
1394: imark = i;
1395: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1396: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1397: }
1398: if (idx) {
1399: *idx = idx_p = mat->rowindices;
1400: if (imark > -1) {
1401: for (i=0; i<imark; i++) {
1402: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1403: }
1404: } else {
1405: for (i=0; i<nzB; i++) {
1406: if (cmap[cworkB[i]/bs] < cstart)
1407: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1408: else break;
1409: }
1410: imark = i;
1411: }
1412: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i];
1413: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1414: }
1415: } else {
1416: if (idx) *idx = 0;
1417: if (v) *v = 0;
1418: }
1419: }
1420: *nz = nztot;
1421: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1422: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1423: return(0);
1424: }
1428: int MatRestoreRow_MPIBAIJ(Mat mat,int row,int *nz,int **idx,PetscScalar **v)
1429: {
1430: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1433: if (baij->getrowactive == PETSC_FALSE) {
1434: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1435: }
1436: baij->getrowactive = PETSC_FALSE;
1437: return(0);
1438: }
1442: int MatGetBlockSize_MPIBAIJ(Mat mat,int *bs)
1443: {
1444: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1447: *bs = baij->bs;
1448: return(0);
1449: }
1453: int MatZeroEntries_MPIBAIJ(Mat A)
1454: {
1455: Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data;
1456: int ierr;
1459: MatZeroEntries(l->A);
1460: MatZeroEntries(l->B);
1461: return(0);
1462: }
1466: int MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1467: {
1468: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data;
1469: Mat A = a->A,B = a->B;
1470: int ierr;
1471: PetscReal isend[5],irecv[5];
1474: info->block_size = (PetscReal)a->bs;
1475: MatGetInfo(A,MAT_LOCAL,info);
1476: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1477: isend[3] = info->memory; isend[4] = info->mallocs;
1478: MatGetInfo(B,MAT_LOCAL,info);
1479: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1480: isend[3] += info->memory; isend[4] += info->mallocs;
1481: if (flag == MAT_LOCAL) {
1482: info->nz_used = isend[0];
1483: info->nz_allocated = isend[1];
1484: info->nz_unneeded = isend[2];
1485: info->memory = isend[3];
1486: info->mallocs = isend[4];
1487: } else if (flag == MAT_GLOBAL_MAX) {
1488: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);
1489: info->nz_used = irecv[0];
1490: info->nz_allocated = irecv[1];
1491: info->nz_unneeded = irecv[2];
1492: info->memory = irecv[3];
1493: info->mallocs = irecv[4];
1494: } else if (flag == MAT_GLOBAL_SUM) {
1495: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);
1496: info->nz_used = irecv[0];
1497: info->nz_allocated = irecv[1];
1498: info->nz_unneeded = irecv[2];
1499: info->memory = irecv[3];
1500: info->mallocs = irecv[4];
1501: } else {
1502: SETERRQ1(1,"Unknown MatInfoType argument %d",flag);
1503: }
1504: info->rows_global = (PetscReal)A->M;
1505: info->columns_global = (PetscReal)A->N;
1506: info->rows_local = (PetscReal)A->m;
1507: info->columns_local = (PetscReal)A->N;
1508: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1509: info->fill_ratio_needed = 0;
1510: info->factor_mallocs = 0;
1511: return(0);
1512: }
1516: int MatSetOption_MPIBAIJ(Mat A,MatOption op)
1517: {
1518: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1519: int ierr;
1522: switch (op) {
1523: case MAT_NO_NEW_NONZERO_LOCATIONS:
1524: case MAT_YES_NEW_NONZERO_LOCATIONS:
1525: case MAT_COLUMNS_UNSORTED:
1526: case MAT_COLUMNS_SORTED:
1527: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1528: case MAT_KEEP_ZEROED_ROWS:
1529: case MAT_NEW_NONZERO_LOCATION_ERR:
1530: MatSetOption(a->A,op);
1531: MatSetOption(a->B,op);
1532: break;
1533: case MAT_ROW_ORIENTED:
1534: a->roworiented = PETSC_TRUE;
1535: MatSetOption(a->A,op);
1536: MatSetOption(a->B,op);
1537: break;
1538: case MAT_ROWS_SORTED:
1539: case MAT_ROWS_UNSORTED:
1540: case MAT_YES_NEW_DIAGONALS:
1541: PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n");
1542: break;
1543: case MAT_COLUMN_ORIENTED:
1544: a->roworiented = PETSC_FALSE;
1545: MatSetOption(a->A,op);
1546: MatSetOption(a->B,op);
1547: break;
1548: case MAT_IGNORE_OFF_PROC_ENTRIES:
1549: a->donotstash = PETSC_TRUE;
1550: break;
1551: case MAT_NO_NEW_DIAGONALS:
1552: SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1553: case MAT_USE_HASH_TABLE:
1554: a->ht_flag = PETSC_TRUE;
1555: break;
1556: default:
1557: SETERRQ(PETSC_ERR_SUP,"unknown option");
1558: }
1559: return(0);
1560: }
1564: int MatTranspose_MPIBAIJ(Mat A,Mat *matout)
1565: {
1566: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data;
1567: Mat_SeqBAIJ *Aloc;
1568: Mat B;
1569: int ierr,M=A->M,N=A->N,*ai,*aj,i,*rvals,j,k,col;
1570: int bs=baij->bs,mbs=baij->mbs;
1571: MatScalar *a;
1572:
1574: if (!matout && M != N) SETERRQ(PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1575: MatCreateMPIBAIJ(A->comm,baij->bs,A->n,A->m,N,M,0,PETSC_NULL,0,PETSC_NULL,&B);
1576:
1577: /* copy over the A part */
1578: Aloc = (Mat_SeqBAIJ*)baij->A->data;
1579: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1580: PetscMalloc(bs*sizeof(int),&rvals);
1581:
1582: for (i=0; i<mbs; i++) {
1583: rvals[0] = bs*(baij->rstart + i);
1584: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
1585: for (j=ai[i]; j<ai[i+1]; j++) {
1586: col = (baij->cstart+aj[j])*bs;
1587: for (k=0; k<bs; k++) {
1588: MatSetValues_MPIBAIJ_MatScalar(B,1,&col,bs,rvals,a,INSERT_VALUES);
1589: col++; a += bs;
1590: }
1591: }
1592: }
1593: /* copy over the B part */
1594: Aloc = (Mat_SeqBAIJ*)baij->B->data;
1595: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1596: for (i=0; i<mbs; i++) {
1597: rvals[0] = bs*(baij->rstart + i);
1598: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
1599: for (j=ai[i]; j<ai[i+1]; j++) {
1600: col = baij->garray[aj[j]]*bs;
1601: for (k=0; k<bs; k++) {
1602: MatSetValues_MPIBAIJ_MatScalar(B,1,&col,bs,rvals,a,INSERT_VALUES);
1603: col++; a += bs;
1604: }
1605: }
1606: }
1607: PetscFree(rvals);
1608: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1609: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
1610:
1611: if (matout) {
1612: *matout = B;
1613: } else {
1614: MatHeaderCopy(A,B);
1615: }
1616: return(0);
1617: }
1621: int MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr)
1622: {
1623: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1624: Mat a = baij->A,b = baij->B;
1625: int ierr,s1,s2,s3;
1628: MatGetLocalSize(mat,&s2,&s3);
1629: if (rr) {
1630: VecGetLocalSize(rr,&s1);
1631: if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1632: /* Overlap communication with computation. */
1633: VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1634: }
1635: if (ll) {
1636: VecGetLocalSize(ll,&s1);
1637: if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1638: (*b->ops->diagonalscale)(b,ll,PETSC_NULL);
1639: }
1640: /* scale the diagonal block */
1641: (*a->ops->diagonalscale)(a,ll,rr);
1643: if (rr) {
1644: /* Do a scatter end and then right scale the off-diagonal block */
1645: VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1646: (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);
1647: }
1648:
1649: return(0);
1650: }
1654: int MatZeroRows_MPIBAIJ(Mat A,IS is,const PetscScalar *diag)
1655: {
1656: Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data;
1657: int i,ierr,N,*rows,*owners = l->rowners,size = l->size;
1658: int *nprocs,j,idx,nsends,row;
1659: int nmax,*svalues,*starts,*owner,nrecvs,rank = l->rank;
1660: int *rvalues,tag = A->tag,count,base,slen,n,*source;
1661: int *lens,imdex,*lrows,*values,bs=l->bs,rstart_bs=l->rstart_bs;
1662: MPI_Comm comm = A->comm;
1663: MPI_Request *send_waits,*recv_waits;
1664: MPI_Status recv_status,*send_status;
1665: IS istmp;
1666: PetscTruth found;
1667:
1669: ISGetLocalSize(is,&N);
1670: ISGetIndices(is,&rows);
1671:
1672: /* first count number of contributors to each processor */
1673: PetscMalloc(2*size*sizeof(int),&nprocs);
1674: PetscMemzero(nprocs,2*size*sizeof(int));
1675: PetscMalloc((N+1)*sizeof(int),&owner); /* see note*/
1676: for (i=0; i<N; i++) {
1677: idx = rows[i];
1678: found = PETSC_FALSE;
1679: for (j=0; j<size; j++) {
1680: if (idx >= owners[j]*bs && idx < owners[j+1]*bs) {
1681: nprocs[2*j]++; nprocs[2*j+1] = 1; owner[i] = j; found = PETSC_TRUE; break;
1682: }
1683: }
1684: if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
1685: }
1686: nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}
1687:
1688: /* inform other processors of number of messages and max length*/
1689: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
1690:
1691: /* post receives: */
1692: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(int),&rvalues);
1693: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
1694: for (i=0; i<nrecvs; i++) {
1695: MPI_Irecv(rvalues+nmax*i,nmax,MPI_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
1696: }
1697:
1698: /* do sends:
1699: 1) starts[i] gives the starting index in svalues for stuff going to
1700: the ith processor
1701: */
1702: PetscMalloc((N+1)*sizeof(int),&svalues);
1703: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
1704: PetscMalloc((size+1)*sizeof(int),&starts);
1705: starts[0] = 0;
1706: for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
1707: for (i=0; i<N; i++) {
1708: svalues[starts[owner[i]]++] = rows[i];
1709: }
1710: ISRestoreIndices(is,&rows);
1711:
1712: starts[0] = 0;
1713: for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
1714: count = 0;
1715: for (i=0; i<size; i++) {
1716: if (nprocs[2*i+1]) {
1717: MPI_Isend(svalues+starts[i],nprocs[2*i],MPI_INT,i,tag,comm,send_waits+count++);
1718: }
1719: }
1720: PetscFree(starts);
1722: base = owners[rank]*bs;
1723:
1724: /* wait on receives */
1725: PetscMalloc(2*(nrecvs+1)*sizeof(int),&lens);
1726: source = lens + nrecvs;
1727: count = nrecvs; slen = 0;
1728: while (count) {
1729: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
1730: /* unpack receives into our local space */
1731: MPI_Get_count(&recv_status,MPI_INT,&n);
1732: source[imdex] = recv_status.MPI_SOURCE;
1733: lens[imdex] = n;
1734: slen += n;
1735: count--;
1736: }
1737: PetscFree(recv_waits);
1738:
1739: /* move the data into the send scatter */
1740: PetscMalloc((slen+1)*sizeof(int),&lrows);
1741: count = 0;
1742: for (i=0; i<nrecvs; i++) {
1743: values = rvalues + i*nmax;
1744: for (j=0; j<lens[i]; j++) {
1745: lrows[count++] = values[j] - base;
1746: }
1747: }
1748: PetscFree(rvalues);
1749: PetscFree(lens);
1750: PetscFree(owner);
1751: PetscFree(nprocs);
1752:
1753: /* actually zap the local rows */
1754: ISCreateGeneral(PETSC_COMM_SELF,slen,lrows,&istmp);
1755: PetscLogObjectParent(A,istmp);
1757: /*
1758: Zero the required rows. If the "diagonal block" of the matrix
1759: is square and the user wishes to set the diagonal we use seperate
1760: code so that MatSetValues() is not called for each diagonal allocating
1761: new memory, thus calling lots of mallocs and slowing things down.
1763: Contributed by: Mathew Knepley
1764: */
1765: /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
1766: MatZeroRows_SeqBAIJ(l->B,istmp,0);
1767: if (diag && (l->A->M == l->A->N)) {
1768: MatZeroRows_SeqBAIJ(l->A,istmp,diag);
1769: } else if (diag) {
1770: MatZeroRows_SeqBAIJ(l->A,istmp,0);
1771: if (((Mat_SeqBAIJ*)l->A->data)->nonew) {
1772: SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\
1773: MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
1774: }
1775: for (i=0; i<slen; i++) {
1776: row = lrows[i] + rstart_bs;
1777: MatSetValues(A,1,&row,1,&row,diag,INSERT_VALUES);
1778: }
1779: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1780: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1781: } else {
1782: MatZeroRows_SeqBAIJ(l->A,istmp,0);
1783: }
1785: ISDestroy(istmp);
1786: PetscFree(lrows);
1788: /* wait on sends */
1789: if (nsends) {
1790: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
1791: MPI_Waitall(nsends,send_waits,send_status);
1792: PetscFree(send_status);
1793: }
1794: PetscFree(send_waits);
1795: PetscFree(svalues);
1797: return(0);
1798: }
1802: int MatPrintHelp_MPIBAIJ(Mat A)
1803: {
1804: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1805: MPI_Comm comm = A->comm;
1806: static int called = 0;
1807: int ierr;
1810: if (!a->rank) {
1811: MatPrintHelp_SeqBAIJ(a->A);
1812: }
1813: if (called) {return(0);} else called = 1;
1814: (*PetscHelpPrintf)(comm," Options for MATMPIBAIJ matrix format (the defaults):\n");
1815: (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");
1816: return(0);
1817: }
1821: int MatSetUnfactored_MPIBAIJ(Mat A)
1822: {
1823: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1824: int ierr;
1827: MatSetUnfactored(a->A);
1828: return(0);
1829: }
1831: static int MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat *);
1835: int MatEqual_MPIBAIJ(Mat A,Mat B,PetscTruth *flag)
1836: {
1837: Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data;
1838: Mat a,b,c,d;
1839: PetscTruth flg;
1840: int ierr;
1843: a = matA->A; b = matA->B;
1844: c = matB->A; d = matB->B;
1846: MatEqual(a,c,&flg);
1847: if (flg == PETSC_TRUE) {
1848: MatEqual(b,d,&flg);
1849: }
1850: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);
1851: return(0);
1852: }
1857: int MatSetUpPreallocation_MPIBAIJ(Mat A)
1858: {
1859: int ierr;
1862: MatMPIBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1863: return(0);
1864: }
1866: /* -------------------------------------------------------------------*/
1867: static struct _MatOps MatOps_Values = {
1868: MatSetValues_MPIBAIJ,
1869: MatGetRow_MPIBAIJ,
1870: MatRestoreRow_MPIBAIJ,
1871: MatMult_MPIBAIJ,
1872: /* 4*/ MatMultAdd_MPIBAIJ,
1873: MatMultTranspose_MPIBAIJ,
1874: MatMultTransposeAdd_MPIBAIJ,
1875: 0,
1876: 0,
1877: 0,
1878: /*10*/ 0,
1879: 0,
1880: 0,
1881: 0,
1882: MatTranspose_MPIBAIJ,
1883: /*15*/ MatGetInfo_MPIBAIJ,
1884: MatEqual_MPIBAIJ,
1885: MatGetDiagonal_MPIBAIJ,
1886: MatDiagonalScale_MPIBAIJ,
1887: MatNorm_MPIBAIJ,
1888: /*20*/ MatAssemblyBegin_MPIBAIJ,
1889: MatAssemblyEnd_MPIBAIJ,
1890: 0,
1891: MatSetOption_MPIBAIJ,
1892: MatZeroEntries_MPIBAIJ,
1893: /*25*/ MatZeroRows_MPIBAIJ,
1894: 0,
1895: 0,
1896: 0,
1897: 0,
1898: /*30*/ MatSetUpPreallocation_MPIBAIJ,
1899: 0,
1900: 0,
1901: 0,
1902: 0,
1903: /*35*/ MatDuplicate_MPIBAIJ,
1904: 0,
1905: 0,
1906: 0,
1907: 0,
1908: /*40*/ 0,
1909: MatGetSubMatrices_MPIBAIJ,
1910: MatIncreaseOverlap_MPIBAIJ,
1911: MatGetValues_MPIBAIJ,
1912: 0,
1913: /*45*/ MatPrintHelp_MPIBAIJ,
1914: MatScale_MPIBAIJ,
1915: 0,
1916: 0,
1917: 0,
1918: /*50*/ MatGetBlockSize_MPIBAIJ,
1919: 0,
1920: 0,
1921: 0,
1922: 0,
1923: /*55*/ 0,
1924: 0,
1925: MatSetUnfactored_MPIBAIJ,
1926: 0,
1927: MatSetValuesBlocked_MPIBAIJ,
1928: /*60*/ 0,
1929: MatDestroy_MPIBAIJ,
1930: MatView_MPIBAIJ,
1931: MatGetPetscMaps_Petsc,
1932: 0,
1933: /*65*/ 0,
1934: 0,
1935: 0,
1936: 0,
1937: 0,
1938: /*70*/ MatGetRowMax_MPIBAIJ,
1939: 0,
1940: 0,
1941: 0,
1942: 0,
1943: /*75*/ 0,
1944: 0,
1945: 0,
1946: 0,
1947: 0,
1948: /*80*/ 0,
1949: 0,
1950: 0,
1951: 0,
1952: 0,
1953: /*85*/ MatLoad_MPIBAIJ
1954: };
1957: EXTERN_C_BEGIN
1960: int MatGetDiagonalBlock_MPIBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a)
1961: {
1963: *a = ((Mat_MPIBAIJ *)A->data)->A;
1964: *iscopy = PETSC_FALSE;
1965: return(0);
1966: }
1967: EXTERN_C_END
1969: EXTERN_C_BEGIN
1972: int MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,int bs,int d_nz,int *d_nnz,int o_nz,int *o_nnz)
1973: {
1974: Mat_MPIBAIJ *b;
1975: int ierr,i;
1978: B->preallocated = PETSC_TRUE;
1979: PetscOptionsGetInt(PETSC_NULL,"-mat_block_size",&bs,PETSC_NULL);
1981: if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive");
1982: if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
1983: if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
1984: if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %d",d_nz);
1985: if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %d",o_nz);
1986: if (d_nnz) {
1987: for (i=0; i<B->m/bs; i++) {
1988: if (d_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %d value %d",i,d_nnz[i]);
1989: }
1990: }
1991: if (o_nnz) {
1992: for (i=0; i<B->m/bs; i++) {
1993: if (o_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %d value %d",i,o_nnz[i]);
1994: }
1995: }
1996:
1997: PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);
1998: PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);
1999: PetscMapCreateMPI(B->comm,B->m,B->M,&B->rmap);
2000: PetscMapCreateMPI(B->comm,B->n,B->N,&B->cmap);
2002: b = (Mat_MPIBAIJ*)B->data;
2003: b->bs = bs;
2004: b->bs2 = bs*bs;
2005: b->mbs = B->m/bs;
2006: b->nbs = B->n/bs;
2007: b->Mbs = B->M/bs;
2008: b->Nbs = B->N/bs;
2010: MPI_Allgather(&b->mbs,1,MPI_INT,b->rowners+1,1,MPI_INT,B->comm);
2011: b->rowners[0] = 0;
2012: for (i=2; i<=b->size; i++) {
2013: b->rowners[i] += b->rowners[i-1];
2014: }
2015: b->rstart = b->rowners[b->rank];
2016: b->rend = b->rowners[b->rank+1];
2018: MPI_Allgather(&b->nbs,1,MPI_INT,b->cowners+1,1,MPI_INT,B->comm);
2019: b->cowners[0] = 0;
2020: for (i=2; i<=b->size; i++) {
2021: b->cowners[i] += b->cowners[i-1];
2022: }
2023: b->cstart = b->cowners[b->rank];
2024: b->cend = b->cowners[b->rank+1];
2026: for (i=0; i<=b->size; i++) {
2027: b->rowners_bs[i] = b->rowners[i]*bs;
2028: }
2029: b->rstart_bs = b->rstart*bs;
2030: b->rend_bs = b->rend*bs;
2031: b->cstart_bs = b->cstart*bs;
2032: b->cend_bs = b->cend*bs;
2034: MatCreateSeqBAIJ(PETSC_COMM_SELF,bs,B->m,B->n,d_nz,d_nnz,&b->A);
2035: PetscLogObjectParent(B,b->A);
2036: MatCreateSeqBAIJ(PETSC_COMM_SELF,bs,B->m,B->N,o_nz,o_nnz,&b->B);
2037: PetscLogObjectParent(B,b->B);
2038: MatStashCreate_Private(B->comm,bs,&B->bstash);
2040: return(0);
2041: }
2042: EXTERN_C_END
2044: EXTERN_C_BEGIN
2045: EXTERN int MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec);
2046: EXTERN int MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal);
2047: EXTERN_C_END
2049: /*MC
2050: MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices.
2052: Options Database Keys:
2053: . -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions()
2055: Level: beginner
2057: .seealso: MatCreateMPIBAIJ
2058: M*/
2060: EXTERN_C_BEGIN
2063: int MatCreate_MPIBAIJ(Mat B)
2064: {
2065: Mat_MPIBAIJ *b;
2066: int ierr;
2067: PetscTruth flg;
2071: PetscNew(Mat_MPIBAIJ,&b);
2072: B->data = (void*)b;
2074: PetscMemzero(b,sizeof(Mat_MPIBAIJ));
2075: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
2076: B->mapping = 0;
2077: B->factor = 0;
2078: B->assembled = PETSC_FALSE;
2080: B->insertmode = NOT_SET_VALUES;
2081: MPI_Comm_rank(B->comm,&b->rank);
2082: MPI_Comm_size(B->comm,&b->size);
2084: /* build local table of row and column ownerships */
2085: PetscMalloc(3*(b->size+2)*sizeof(int),&b->rowners);
2086: PetscLogObjectMemory(B,3*(b->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPIBAIJ));
2087: b->cowners = b->rowners + b->size + 2;
2088: b->rowners_bs = b->cowners + b->size + 2;
2090: /* build cache for off array entries formed */
2091: MatStashCreate_Private(B->comm,1,&B->stash);
2092: b->donotstash = PETSC_FALSE;
2093: b->colmap = PETSC_NULL;
2094: b->garray = PETSC_NULL;
2095: b->roworiented = PETSC_TRUE;
2097: #if defined(PETSC_USE_MAT_SINGLE)
2098: /* stuff for MatSetValues_XXX in single precision */
2099: b->setvalueslen = 0;
2100: b->setvaluescopy = PETSC_NULL;
2101: #endif
2103: /* stuff used in block assembly */
2104: b->barray = 0;
2106: /* stuff used for matrix vector multiply */
2107: b->lvec = 0;
2108: b->Mvctx = 0;
2110: /* stuff for MatGetRow() */
2111: b->rowindices = 0;
2112: b->rowvalues = 0;
2113: b->getrowactive = PETSC_FALSE;
2115: /* hash table stuff */
2116: b->ht = 0;
2117: b->hd = 0;
2118: b->ht_size = 0;
2119: b->ht_flag = PETSC_FALSE;
2120: b->ht_fact = 0;
2121: b->ht_total_ct = 0;
2122: b->ht_insert_ct = 0;
2124: PetscOptionsHasName(PETSC_NULL,"-mat_use_hash_table",&flg);
2125: if (flg) {
2126: PetscReal fact = 1.39;
2127: MatSetOption(B,MAT_USE_HASH_TABLE);
2128: PetscOptionsGetReal(PETSC_NULL,"-mat_use_hash_table",&fact,PETSC_NULL);
2129: if (fact <= 1.0) fact = 1.39;
2130: MatMPIBAIJSetHashTableFactor(B,fact);
2131: PetscLogInfo(0,"MatCreateMPIBAIJ:Hash table Factor used %5.2f\n",fact);
2132: }
2133: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
2134: "MatStoreValues_MPIBAIJ",
2135: MatStoreValues_MPIBAIJ);
2136: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
2137: "MatRetrieveValues_MPIBAIJ",
2138: MatRetrieveValues_MPIBAIJ);
2139: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
2140: "MatGetDiagonalBlock_MPIBAIJ",
2141: MatGetDiagonalBlock_MPIBAIJ);
2142: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIBAIJSetPreallocation_C",
2143: "MatMPIBAIJSetPreallocation_MPIBAIJ",
2144: MatMPIBAIJSetPreallocation_MPIBAIJ);
2145: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
2146: "MatDiagonalScaleLocal_MPIBAIJ",
2147: MatDiagonalScaleLocal_MPIBAIJ);
2148: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatSetHashTableFactor_C",
2149: "MatSetHashTableFactor_MPIBAIJ",
2150: MatSetHashTableFactor_MPIBAIJ);
2151: return(0);
2152: }
2153: EXTERN_C_END
2155: /*MC
2156: MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices.
2158: This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator,
2159: and MATMPIBAIJ otherwise.
2161: Options Database Keys:
2162: . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions()
2164: Level: beginner
2166: .seealso: MatCreateMPIBAIJ,MATSEQBAIJ,MATMPIBAIJ
2167: M*/
2169: EXTERN_C_BEGIN
2172: int MatCreate_BAIJ(Mat A) {
2173: int ierr,size;
2176: PetscObjectChangeTypeName((PetscObject)A,MATBAIJ);
2177: MPI_Comm_size(A->comm,&size);
2178: if (size == 1) {
2179: MatSetType(A,MATSEQBAIJ);
2180: } else {
2181: MatSetType(A,MATMPIBAIJ);
2182: }
2183: return(0);
2184: }
2185: EXTERN_C_END
2189: /*@C
2190: MatMPIBAIJSetPreallocation - Creates a sparse parallel matrix in block AIJ format
2191: (block compressed row). For good matrix assembly performance
2192: the user should preallocate the matrix storage by setting the parameters
2193: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
2194: performance can be increased by more than a factor of 50.
2196: Collective on Mat
2198: Input Parameters:
2199: + A - the matrix
2200: . bs - size of blockk
2201: . d_nz - number of block nonzeros per block row in diagonal portion of local
2202: submatrix (same for all local rows)
2203: . d_nnz - array containing the number of block nonzeros in the various block rows
2204: of the in diagonal portion of the local (possibly different for each block
2205: row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero.
2206: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
2207: submatrix (same for all local rows).
2208: - o_nnz - array containing the number of nonzeros in the various block rows of the
2209: off-diagonal portion of the local submatrix (possibly different for
2210: each block row) or PETSC_NULL.
2212: Output Parameter:
2215: Options Database Keys:
2216: . -mat_no_unroll - uses code that does not unroll the loops in the
2217: block calculations (much slower)
2218: . -mat_block_size - size of the blocks to use
2220: Notes:
2221: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
2222: than it must be used on all processors that share the object for that argument.
2224: Storage Information:
2225: For a square global matrix we define each processor's diagonal portion
2226: to be its local rows and the corresponding columns (a square submatrix);
2227: each processor's off-diagonal portion encompasses the remainder of the
2228: local matrix (a rectangular submatrix).
2230: The user can specify preallocated storage for the diagonal part of
2231: the local submatrix with either d_nz or d_nnz (not both). Set
2232: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
2233: memory allocation. Likewise, specify preallocated storage for the
2234: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
2236: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
2237: the figure below we depict these three local rows and all columns (0-11).
2239: .vb
2240: 0 1 2 3 4 5 6 7 8 9 10 11
2241: -------------------
2242: row 3 | o o o d d d o o o o o o
2243: row 4 | o o o d d d o o o o o o
2244: row 5 | o o o d d d o o o o o o
2245: -------------------
2246: .ve
2247:
2248: Thus, any entries in the d locations are stored in the d (diagonal)
2249: submatrix, and any entries in the o locations are stored in the
2250: o (off-diagonal) submatrix. Note that the d and the o submatrices are
2251: stored simply in the MATSEQBAIJ format for compressed row storage.
2253: Now d_nz should indicate the number of block nonzeros per row in the d matrix,
2254: and o_nz should indicate the number of block nonzeros per row in the o matrix.
2255: In general, for PDE problems in which most nonzeros are near the diagonal,
2256: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
2257: or you will get TERRIBLE performance; see the users' manual chapter on
2258: matrices.
2260: Level: intermediate
2262: .keywords: matrix, block, aij, compressed row, sparse, parallel
2264: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
2265: @*/
2266: int MatMPIBAIJSetPreallocation(Mat B,int bs,int d_nz,const int d_nnz[],int o_nz,const int o_nnz[])
2267: {
2268: int ierr,(*f)(Mat,int,int,const int[],int,const int[]);
2271: PetscObjectQueryFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",(void (**)(void))&f);
2272: if (f) {
2273: (*f)(B,bs,d_nz,d_nnz,o_nz,o_nnz);
2274: }
2275: return(0);
2276: }
2280: /*@C
2281: MatCreateMPIBAIJ - Creates a sparse parallel matrix in block AIJ format
2282: (block compressed row). For good matrix assembly performance
2283: the user should preallocate the matrix storage by setting the parameters
2284: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
2285: performance can be increased by more than a factor of 50.
2287: Collective on MPI_Comm
2289: Input Parameters:
2290: + comm - MPI communicator
2291: . bs - size of blockk
2292: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
2293: This value should be the same as the local size used in creating the
2294: y vector for the matrix-vector product y = Ax.
2295: . n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
2296: This value should be the same as the local size used in creating the
2297: x vector for the matrix-vector product y = Ax.
2298: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
2299: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
2300: . d_nz - number of nonzero blocks per block row in diagonal portion of local
2301: submatrix (same for all local rows)
2302: . d_nnz - array containing the number of nonzero blocks in the various block rows
2303: of the in diagonal portion of the local (possibly different for each block
2304: row) or PETSC_NULL. You must leave room for the diagonal entry even if it is zero.
2305: . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local
2306: submatrix (same for all local rows).
2307: - o_nnz - array containing the number of nonzero blocks in the various block rows of the
2308: off-diagonal portion of the local submatrix (possibly different for
2309: each block row) or PETSC_NULL.
2311: Output Parameter:
2312: . A - the matrix
2314: Options Database Keys:
2315: . -mat_no_unroll - uses code that does not unroll the loops in the
2316: block calculations (much slower)
2317: . -mat_block_size - size of the blocks to use
2319: Notes:
2320: A nonzero block is any block that as 1 or more nonzeros in it
2322: The user MUST specify either the local or global matrix dimensions
2323: (possibly both).
2325: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
2326: than it must be used on all processors that share the object for that argument.
2328: Storage Information:
2329: For a square global matrix we define each processor's diagonal portion
2330: to be its local rows and the corresponding columns (a square submatrix);
2331: each processor's off-diagonal portion encompasses the remainder of the
2332: local matrix (a rectangular submatrix).
2334: The user can specify preallocated storage for the diagonal part of
2335: the local submatrix with either d_nz or d_nnz (not both). Set
2336: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
2337: memory allocation. Likewise, specify preallocated storage for the
2338: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
2340: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
2341: the figure below we depict these three local rows and all columns (0-11).
2343: .vb
2344: 0 1 2 3 4 5 6 7 8 9 10 11
2345: -------------------
2346: row 3 | o o o d d d o o o o o o
2347: row 4 | o o o d d d o o o o o o
2348: row 5 | o o o d d d o o o o o o
2349: -------------------
2350: .ve
2351:
2352: Thus, any entries in the d locations are stored in the d (diagonal)
2353: submatrix, and any entries in the o locations are stored in the
2354: o (off-diagonal) submatrix. Note that the d and the o submatrices are
2355: stored simply in the MATSEQBAIJ format for compressed row storage.
2357: Now d_nz should indicate the number of block nonzeros per row in the d matrix,
2358: and o_nz should indicate the number of block nonzeros per row in the o matrix.
2359: In general, for PDE problems in which most nonzeros are near the diagonal,
2360: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
2361: or you will get TERRIBLE performance; see the users' manual chapter on
2362: matrices.
2364: Level: intermediate
2366: .keywords: matrix, block, aij, compressed row, sparse, parallel
2368: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
2369: @*/
2370: int MatCreateMPIBAIJ(MPI_Comm comm,int bs,int m,int n,int M,int N,int d_nz,const int d_nnz[],int o_nz,const int o_nnz[],Mat *A)
2371: {
2372: int ierr,size;
2375: MatCreate(comm,m,n,M,N,A);
2376: MPI_Comm_size(comm,&size);
2377: if (size > 1) {
2378: MatSetType(*A,MATMPIBAIJ);
2379: MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
2380: } else {
2381: MatSetType(*A,MATSEQBAIJ);
2382: MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
2383: }
2384: return(0);
2385: }
2389: static int MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2390: {
2391: Mat mat;
2392: Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data;
2393: int ierr,len=0;
2396: *newmat = 0;
2397: MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);
2398: MatSetType(mat,MATMPIBAIJ);
2399: mat->preallocated = PETSC_TRUE;
2400: mat->assembled = PETSC_TRUE;
2401: a = (Mat_MPIBAIJ*)mat->data;
2402: a->bs = oldmat->bs;
2403: a->bs2 = oldmat->bs2;
2404: a->mbs = oldmat->mbs;
2405: a->nbs = oldmat->nbs;
2406: a->Mbs = oldmat->Mbs;
2407: a->Nbs = oldmat->Nbs;
2408:
2409: a->rstart = oldmat->rstart;
2410: a->rend = oldmat->rend;
2411: a->cstart = oldmat->cstart;
2412: a->cend = oldmat->cend;
2413: a->size = oldmat->size;
2414: a->rank = oldmat->rank;
2415: a->donotstash = oldmat->donotstash;
2416: a->roworiented = oldmat->roworiented;
2417: a->rowindices = 0;
2418: a->rowvalues = 0;
2419: a->getrowactive = PETSC_FALSE;
2420: a->barray = 0;
2421: a->rstart_bs = oldmat->rstart_bs;
2422: a->rend_bs = oldmat->rend_bs;
2423: a->cstart_bs = oldmat->cstart_bs;
2424: a->cend_bs = oldmat->cend_bs;
2426: /* hash table stuff */
2427: a->ht = 0;
2428: a->hd = 0;
2429: a->ht_size = 0;
2430: a->ht_flag = oldmat->ht_flag;
2431: a->ht_fact = oldmat->ht_fact;
2432: a->ht_total_ct = 0;
2433: a->ht_insert_ct = 0;
2435: PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(int));
2436: MatStashCreate_Private(matin->comm,1,&mat->stash);
2437: MatStashCreate_Private(matin->comm,oldmat->bs,&mat->bstash);
2438: if (oldmat->colmap) {
2439: #if defined (PETSC_USE_CTABLE)
2440: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2441: #else
2442: PetscMalloc((a->Nbs)*sizeof(int),&a->colmap);
2443: PetscLogObjectMemory(mat,(a->Nbs)*sizeof(int));
2444: PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(int));
2445: #endif
2446: } else a->colmap = 0;
2447: if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
2448: PetscMalloc(len*sizeof(int),&a->garray);
2449: PetscLogObjectMemory(mat,len*sizeof(int));
2450: PetscMemcpy(a->garray,oldmat->garray,len*sizeof(int));
2451: } else a->garray = 0;
2452:
2453: VecDuplicate(oldmat->lvec,&a->lvec);
2454: PetscLogObjectParent(mat,a->lvec);
2455: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2457: PetscLogObjectParent(mat,a->Mvctx);
2458: MatDuplicate(oldmat->A,cpvalues,&a->A);
2459: PetscLogObjectParent(mat,a->A);
2460: MatDuplicate(oldmat->B,cpvalues,&a->B);
2461: PetscLogObjectParent(mat,a->B);
2462: PetscFListDuplicate(matin->qlist,&mat->qlist);
2463: *newmat = mat;
2464: return(0);
2465: }
2467: #include petscsys.h
2471: int MatLoad_MPIBAIJ(PetscViewer viewer,MatType type,Mat *newmat)
2472: {
2473: Mat A;
2474: int i,nz,ierr,j,rstart,rend,fd;
2475: PetscScalar *vals,*buf;
2476: MPI_Comm comm = ((PetscObject)viewer)->comm;
2477: MPI_Status status;
2478: int header[4],rank,size,*rowlengths = 0,M,N,m,*rowners,*browners,maxnz,*cols;
2479: int *locrowlens,*sndcounts = 0,*procsnz = 0,jj,*mycols,*ibuf;
2480: int tag = ((PetscObject)viewer)->tag,bs=1,Mbs,mbs,extra_rows;
2481: int *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount;
2482: int dcount,kmax,k,nzcount,tmp;
2483:
2485: PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);
2487: MPI_Comm_size(comm,&size);
2488: MPI_Comm_rank(comm,&rank);
2489: if (!rank) {
2490: PetscViewerBinaryGetDescriptor(viewer,&fd);
2491: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
2492: if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2493: if (header[3] < 0) {
2494: SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPIBAIJ");
2495: }
2496: }
2498: MPI_Bcast(header+1,3,MPI_INT,0,comm);
2499: M = header[1]; N = header[2];
2501: if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices");
2503: /*
2504: This code adds extra rows to make sure the number of rows is
2505: divisible by the blocksize
2506: */
2507: Mbs = M/bs;
2508: extra_rows = bs - M + bs*(Mbs);
2509: if (extra_rows == bs) extra_rows = 0;
2510: else Mbs++;
2511: if (extra_rows &&!rank) {
2512: PetscLogInfo(0,"MatLoad_MPIBAIJ:Padding loaded matrix to match blocksize\n");
2513: }
2515: /* determine ownership of all rows */
2516: mbs = Mbs/size + ((Mbs % size) > rank);
2517: m = mbs*bs;
2518: PetscMalloc(2*(size+2)*sizeof(int),&rowners);
2519: browners = rowners + size + 1;
2520: MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);
2521: rowners[0] = 0;
2522: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
2523: for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
2524: rstart = rowners[rank];
2525: rend = rowners[rank+1];
2527: /* distribute row lengths to all processors */
2528: PetscMalloc((rend-rstart)*bs*sizeof(int),&locrowlens);
2529: if (!rank) {
2530: PetscMalloc((M+extra_rows)*sizeof(int),&rowlengths);
2531: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
2532: for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1;
2533: PetscMalloc(size*sizeof(int),&sndcounts);
2534: for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i];
2535: MPI_Scatterv(rowlengths,sndcounts,browners,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);
2536: PetscFree(sndcounts);
2537: } else {
2538: MPI_Scatterv(0,0,0,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);
2539: }
2541: if (!rank) {
2542: /* calculate the number of nonzeros on each processor */
2543: PetscMalloc(size*sizeof(int),&procsnz);
2544: PetscMemzero(procsnz,size*sizeof(int));
2545: for (i=0; i<size; i++) {
2546: for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) {
2547: procsnz[i] += rowlengths[j];
2548: }
2549: }
2550: PetscFree(rowlengths);
2551:
2552: /* determine max buffer needed and allocate it */
2553: maxnz = 0;
2554: for (i=0; i<size; i++) {
2555: maxnz = PetscMax(maxnz,procsnz[i]);
2556: }
2557: PetscMalloc(maxnz*sizeof(int),&cols);
2559: /* read in my part of the matrix column indices */
2560: nz = procsnz[0];
2561: PetscMalloc(nz*sizeof(int),&ibuf);
2562: mycols = ibuf;
2563: if (size == 1) nz -= extra_rows;
2564: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
2565: if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; }
2567: /* read in every ones (except the last) and ship off */
2568: for (i=1; i<size-1; i++) {
2569: nz = procsnz[i];
2570: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2571: MPI_Send(cols,nz,MPI_INT,i,tag,comm);
2572: }
2573: /* read in the stuff for the last proc */
2574: if (size != 1) {
2575: nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */
2576: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2577: for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
2578: MPI_Send(cols,nz+extra_rows,MPI_INT,size-1,tag,comm);
2579: }
2580: PetscFree(cols);
2581: } else {
2582: /* determine buffer space needed for message */
2583: nz = 0;
2584: for (i=0; i<m; i++) {
2585: nz += locrowlens[i];
2586: }
2587: PetscMalloc(nz*sizeof(int),&ibuf);
2588: mycols = ibuf;
2589: /* receive message of column indices*/
2590: MPI_Recv(mycols,nz,MPI_INT,0,tag,comm,&status);
2591: MPI_Get_count(&status,MPI_INT,&maxnz);
2592: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2593: }
2594:
2595: /* loop over local rows, determining number of off diagonal entries */
2596: PetscMalloc(2*(rend-rstart+1)*sizeof(int),&dlens);
2597: odlens = dlens + (rend-rstart);
2598: PetscMalloc(3*Mbs*sizeof(int),&mask);
2599: PetscMemzero(mask,3*Mbs*sizeof(int));
2600: masked1 = mask + Mbs;
2601: masked2 = masked1 + Mbs;
2602: rowcount = 0; nzcount = 0;
2603: for (i=0; i<mbs; i++) {
2604: dcount = 0;
2605: odcount = 0;
2606: for (j=0; j<bs; j++) {
2607: kmax = locrowlens[rowcount];
2608: for (k=0; k<kmax; k++) {
2609: tmp = mycols[nzcount++]/bs;
2610: if (!mask[tmp]) {
2611: mask[tmp] = 1;
2612: if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp;
2613: else masked1[dcount++] = tmp;
2614: }
2615: }
2616: rowcount++;
2617: }
2618:
2619: dlens[i] = dcount;
2620: odlens[i] = odcount;
2622: /* zero out the mask elements we set */
2623: for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
2624: for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
2625: }
2627: /* create our matrix */
2628: MatCreate(comm,m,m,M+extra_rows,N+extra_rows,&A);
2629: MatSetType(A,type);CHKERRQ(ierr)
2630: MatMPIBAIJSetPreallocation(A,bs,0,dlens,0,odlens);
2632: /* Why doesn't this called using MatSetOption(A,MAT_COLUMNS_SORTED); */
2633: MatSetOption(A,MAT_COLUMNS_SORTED);
2634:
2635: if (!rank) {
2636: PetscMalloc(maxnz*sizeof(PetscScalar),&buf);
2637: /* read in my part of the matrix numerical values */
2638: nz = procsnz[0];
2639: vals = buf;
2640: mycols = ibuf;
2641: if (size == 1) nz -= extra_rows;
2642: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2643: if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; }
2645: /* insert into matrix */
2646: jj = rstart*bs;
2647: for (i=0; i<m; i++) {
2648: MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2649: mycols += locrowlens[i];
2650: vals += locrowlens[i];
2651: jj++;
2652: }
2653: /* read in other processors (except the last one) and ship out */
2654: for (i=1; i<size-1; i++) {
2655: nz = procsnz[i];
2656: vals = buf;
2657: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2658: MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);
2659: }
2660: /* the last proc */
2661: if (size != 1){
2662: nz = procsnz[i] - extra_rows;
2663: vals = buf;
2664: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2665: for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
2666: MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);
2667: }
2668: PetscFree(procsnz);
2669: } else {
2670: /* receive numeric values */
2671: PetscMalloc(nz*sizeof(PetscScalar),&buf);
2673: /* receive message of values*/
2674: vals = buf;
2675: mycols = ibuf;
2676: MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);
2677: MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2678: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2680: /* insert into matrix */
2681: jj = rstart*bs;
2682: for (i=0; i<m; i++) {
2683: MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2684: mycols += locrowlens[i];
2685: vals += locrowlens[i];
2686: jj++;
2687: }
2688: }
2689: PetscFree(locrowlens);
2690: PetscFree(buf);
2691: PetscFree(ibuf);
2692: PetscFree(rowners);
2693: PetscFree(dlens);
2694: PetscFree(mask);
2695: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2696: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2698: *newmat = A;
2699: return(0);
2700: }
2704: /*@
2705: MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.
2707: Input Parameters:
2708: . mat - the matrix
2709: . fact - factor
2711: Collective on Mat
2713: Level: advanced
2715: Notes:
2716: This can also be set by the command line option: -mat_use_hash_table fact
2718: .keywords: matrix, hashtable, factor, HT
2720: .seealso: MatSetOption()
2721: @*/
2722: int MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact)
2723: {
2724: int ierr,(*f)(Mat,PetscReal);
2727: PetscObjectQueryFunction((PetscObject)mat,"MatSetHashTableFactor_C",(void (**)(void))&f);
2728: if (f) {
2729: (*f)(mat,fact);
2730: }
2731: return(0);
2732: }
2736: int MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact)
2737: {
2738: Mat_MPIBAIJ *baij;
2742: baij = (Mat_MPIBAIJ*)mat->data;
2743: baij->ht_fact = fact;
2744: return(0);
2745: }
2749: int MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,int *colmap[])
2750: {
2751: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
2753: *Ad = a->A;
2754: *Ao = a->B;
2755: *colmap = a->garray;
2756: return(0);
2757: }