Actual source code: mpisbaij.c
2: #include src/mat/impls/baij/mpi/mpibaij.h
3: #include mpisbaij.h
4: #include src/mat/impls/sbaij/seq/sbaij.h
6: EXTERN PetscErrorCode MatSetUpMultiply_MPISBAIJ(Mat);
7: EXTERN PetscErrorCode MatSetUpMultiply_MPISBAIJ_2comm(Mat);
8: EXTERN PetscErrorCode DisAssemble_MPISBAIJ(Mat);
9: EXTERN PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat,PetscInt,IS[],PetscInt);
10: EXTERN PetscErrorCode MatGetValues_SeqSBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],PetscScalar []);
11: EXTERN PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],PetscScalar []);
12: EXTERN PetscErrorCode MatSetValues_SeqSBAIJ(Mat,PetscInt,const PetscInt [],PetscInt,const PetscInt [],const PetscScalar [],InsertMode);
13: EXTERN PetscErrorCode MatSetValuesBlocked_SeqSBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
14: EXTERN PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
15: EXTERN PetscErrorCode MatGetRow_SeqSBAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**);
16: EXTERN PetscErrorCode MatRestoreRow_SeqSBAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**);
17: EXTERN PetscErrorCode MatPrintHelp_SeqSBAIJ(Mat);
18: EXTERN PetscErrorCode MatZeroRows_SeqSBAIJ(Mat,IS,PetscScalar*);
19: EXTERN PetscErrorCode MatZeroRows_SeqBAIJ(Mat,IS,PetscScalar *);
20: EXTERN PetscErrorCode MatGetRowMax_MPISBAIJ(Mat,Vec);
21: EXTERN PetscErrorCode MatRelax_MPISBAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec);
23: /* UGLY, ugly, ugly
24: When MatScalar == PetscScalar the function MatSetValuesBlocked_MPIBAIJ_MatScalar() does
25: not exist. Otherwise ..._MatScalar() takes matrix elements in single precision and
26: inserts them into the single precision data structure. The function MatSetValuesBlocked_MPIBAIJ()
27: converts the entries into single precision and then calls ..._MatScalar() to put them
28: into the single precision data structures.
29: */
30: #if defined(PETSC_USE_MAT_SINGLE)
31: EXTERN PetscErrorCode MatSetValuesBlocked_SeqSBAIJ_MatScalar(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const MatScalar[],InsertMode);
32: EXTERN PetscErrorCode MatSetValues_MPISBAIJ_MatScalar(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const MatScalar[],InsertMode);
33: EXTERN PetscErrorCode MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const MatScalar[],InsertMode);
34: EXTERN PetscErrorCode MatSetValues_MPISBAIJ_HT_MatScalar(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const MatScalar[],InsertMode);
35: EXTERN PetscErrorCode MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const MatScalar[],InsertMode);
36: #else
37: #define MatSetValuesBlocked_SeqSBAIJ_MatScalar MatSetValuesBlocked_SeqSBAIJ
38: #define MatSetValues_MPISBAIJ_MatScalar MatSetValues_MPISBAIJ
39: #define MatSetValuesBlocked_MPISBAIJ_MatScalar MatSetValuesBlocked_MPISBAIJ
40: #define MatSetValues_MPISBAIJ_HT_MatScalar MatSetValues_MPISBAIJ_HT
41: #define MatSetValuesBlocked_MPISBAIJ_HT_MatScalar MatSetValuesBlocked_MPISBAIJ_HT
42: #endif
47: PetscErrorCode MatStoreValues_MPISBAIJ(Mat mat)
48: {
49: Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data;
53: MatStoreValues(aij->A);
54: MatStoreValues(aij->B);
55: return(0);
56: }
62: PetscErrorCode MatRetrieveValues_MPISBAIJ(Mat mat)
63: {
64: Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data;
68: MatRetrieveValues(aij->A);
69: MatRetrieveValues(aij->B);
70: return(0);
71: }
75: #define CHUNKSIZE 10
77: #define MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv) \
78: { \
79: \
80: brow = row/bs; \
81: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
82: rmax = aimax[brow]; nrow = ailen[brow]; \
83: bcol = col/bs; \
84: ridx = row % bs; cidx = col % bs; \
85: low = 0; high = nrow; \
86: while (high-low > 3) { \
87: t = (low+high)/2; \
88: if (rp[t] > bcol) high = t; \
89: else low = t; \
90: } \
91: for (_i=low; _i<high; _i++) { \
92: if (rp[_i] > bcol) break; \
93: if (rp[_i] == bcol) { \
94: bap = ap + bs2*_i + bs*cidx + ridx; \
95: if (addv == ADD_VALUES) *bap += value; \
96: else *bap = value; \
97: goto a_noinsert; \
98: } \
99: } \
100: if (a->nonew == 1) goto a_noinsert; \
101: else if (a->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
102: if (nrow >= rmax) { \
103: /* there is no extra room in row, therefore enlarge */ \
104: PetscInt new_nz = ai[a->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
105: MatScalar *new_a; \
106: \
107: if (a->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col); \
108: \
109: /* malloc new storage space */ \
110: len = new_nz*(sizeof(PetscInt)+bs2*sizeof(MatScalar))+(a->mbs+1)*sizeof(PetscInt); \
111: PetscMalloc(len,&new_a); \
112: new_j = (PetscInt*)(new_a + bs2*new_nz); \
113: new_i = new_j + new_nz; \
114: \
115: /* copy over old data into new slots */ \
116: for (ii=0; ii<brow+1; ii++) {new_i[ii] = ai[ii];} \
117: for (ii=brow+1; ii<a->mbs+1; ii++) {new_i[ii] = ai[ii]+CHUNKSIZE;} \
118: PetscMemcpy(new_j,aj,(ai[brow]+nrow)*sizeof(PetscInt)); \
119: len = (new_nz - CHUNKSIZE - ai[brow] - nrow); \
120: PetscMemcpy(new_j+ai[brow]+nrow+CHUNKSIZE,aj+ai[brow]+nrow,len*sizeof(PetscInt)); \
121: PetscMemcpy(new_a,aa,(ai[brow]+nrow)*bs2*sizeof(MatScalar)); \
122: PetscMemzero(new_a+bs2*(ai[brow]+nrow),bs2*CHUNKSIZE*sizeof(PetscScalar)); \
123: PetscMemcpy(new_a+bs2*(ai[brow]+nrow+CHUNKSIZE), \
124: aa+bs2*(ai[brow]+nrow),bs2*len*sizeof(MatScalar)); \
125: /* free up old matrix storage */ \
126: PetscFree(a->a); \
127: if (!a->singlemalloc) { \
128: PetscFree(a->i); \
129: PetscFree(a->j);\
130: } \
131: aa = a->a = new_a; ai = a->i = new_i; aj = a->j = new_j; \
132: a->singlemalloc = PETSC_TRUE; \
133: \
134: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
135: rmax = aimax[brow] = aimax[brow] + CHUNKSIZE; \
136: PetscLogObjectMemory(A,CHUNKSIZE*(sizeof(PetscInt) + bs2*sizeof(MatScalar))); \
137: a->maxnz += bs2*CHUNKSIZE; \
138: a->reallocs++; \
139: a->nz++; \
140: } \
141: N = nrow++ - 1; \
142: /* shift up all the later entries in this row */ \
143: for (ii=N; ii>=_i; ii--) { \
144: rp[ii+1] = rp[ii]; \
145: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
146: } \
147: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); } \
148: rp[_i] = bcol; \
149: ap[bs2*_i + bs*cidx + ridx] = value; \
150: a_noinsert:; \
151: ailen[brow] = nrow; \
152: }
153: #ifndef MatSetValues_SeqBAIJ_B_Private
154: #define MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv) \
155: { \
156: brow = row/bs; \
157: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
158: rmax = bimax[brow]; nrow = bilen[brow]; \
159: bcol = col/bs; \
160: ridx = row % bs; cidx = col % bs; \
161: low = 0; high = nrow; \
162: while (high-low > 3) { \
163: t = (low+high)/2; \
164: if (rp[t] > bcol) high = t; \
165: else low = t; \
166: } \
167: for (_i=low; _i<high; _i++) { \
168: if (rp[_i] > bcol) break; \
169: if (rp[_i] == bcol) { \
170: bap = ap + bs2*_i + bs*cidx + ridx; \
171: if (addv == ADD_VALUES) *bap += value; \
172: else *bap = value; \
173: goto b_noinsert; \
174: } \
175: } \
176: if (b->nonew == 1) goto b_noinsert; \
177: else if (b->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
178: if (nrow >= rmax) { \
179: /* there is no extra room in row, therefore enlarge */ \
180: PetscInt new_nz = bi[b->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
181: MatScalar *new_a; \
182: \
183: if (b->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col); \
184: \
185: /* malloc new storage space */ \
186: len = new_nz*(sizeof(PetscInt)+bs2*sizeof(MatScalar))+(b->mbs+1)*sizeof(PetscInt); \
187: PetscMalloc(len,&new_a); \
188: new_j = (PetscInt*)(new_a + bs2*new_nz); \
189: new_i = new_j + new_nz; \
190: \
191: /* copy over old data into new slots */ \
192: for (ii=0; ii<brow+1; ii++) {new_i[ii] = bi[ii];} \
193: for (ii=brow+1; ii<b->mbs+1; ii++) {new_i[ii] = bi[ii]+CHUNKSIZE;} \
194: PetscMemcpy(new_j,bj,(bi[brow]+nrow)*sizeof(PetscInt)); \
195: len = (new_nz - CHUNKSIZE - bi[brow] - nrow); \
196: PetscMemcpy(new_j+bi[brow]+nrow+CHUNKSIZE,bj+bi[brow]+nrow,len*sizeof(PetscInt)); \
197: PetscMemcpy(new_a,ba,(bi[brow]+nrow)*bs2*sizeof(MatScalar)); \
198: PetscMemzero(new_a+bs2*(bi[brow]+nrow),bs2*CHUNKSIZE*sizeof(MatScalar)); \
199: PetscMemcpy(new_a+bs2*(bi[brow]+nrow+CHUNKSIZE), \
200: ba+bs2*(bi[brow]+nrow),bs2*len*sizeof(MatScalar)); \
201: /* free up old matrix storage */ \
202: PetscFree(b->a); \
203: if (!b->singlemalloc) { \
204: PetscFree(b->i); \
205: PetscFree(b->j); \
206: } \
207: ba = b->a = new_a; bi = b->i = new_i; bj = b->j = new_j; \
208: b->singlemalloc = PETSC_TRUE; \
209: \
210: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
211: rmax = bimax[brow] = bimax[brow] + CHUNKSIZE; \
212: PetscLogObjectMemory(B,CHUNKSIZE*(sizeof(PetscInt) + bs2*sizeof(MatScalar))); \
213: b->maxnz += bs2*CHUNKSIZE; \
214: b->reallocs++; \
215: b->nz++; \
216: } \
217: N = nrow++ - 1; \
218: /* shift up all the later entries in this row */ \
219: for (ii=N; ii>=_i; ii--) { \
220: rp[ii+1] = rp[ii]; \
221: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
222: } \
223: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));} \
224: rp[_i] = bcol; \
225: ap[bs2*_i + bs*cidx + ridx] = value; \
226: b_noinsert:; \
227: bilen[brow] = nrow; \
228: }
229: #endif
231: #if defined(PETSC_USE_MAT_SINGLE)
234: PetscErrorCode MatSetValues_MPISBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
235: {
236: Mat_MPISBAIJ *b = (Mat_MPISBAIJ*)mat->data;
238: PetscInt i,N = m*n;
239: MatScalar *vsingle;
242: if (N > b->setvalueslen) {
243: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
244: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
245: b->setvalueslen = N;
246: }
247: vsingle = b->setvaluescopy;
249: for (i=0; i<N; i++) {
250: vsingle[i] = v[i];
251: }
252: MatSetValues_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);
253: return(0);
254: }
258: PetscErrorCode MatSetValuesBlocked_MPISBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
259: {
260: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
262: PetscInt i,N = m*n*b->bs2;
263: MatScalar *vsingle;
266: if (N > b->setvalueslen) {
267: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
268: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
269: b->setvalueslen = N;
270: }
271: vsingle = b->setvaluescopy;
272: for (i=0; i<N; i++) {
273: vsingle[i] = v[i];
274: }
275: MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);
276: return(0);
277: }
281: PetscErrorCode MatSetValues_MPISBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
282: {
283: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
285: PetscInt i,N = m*n;
286: MatScalar *vsingle;
289: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
290: /* return(0); */
291: }
295: PetscErrorCode MatSetValuesBlocked_MPISBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
296: {
297: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
299: PetscInt i,N = m*n*b->bs2;
300: MatScalar *vsingle;
303: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
304: /* return(0); */
305: }
306: #endif
308: /* Only add/insert a(i,j) with i<=j (blocks).
309: Any a(i,j) with i>j input by user is ingored.
310: */
313: PetscErrorCode MatSetValues_MPISBAIJ_MatScalar(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const MatScalar v[],InsertMode addv)
314: {
315: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
316: MatScalar value;
317: PetscTruth roworiented = baij->roworiented;
319: PetscInt i,j,row,col;
320: PetscInt rstart_orig=baij->rstart_bs;
321: PetscInt rend_orig=baij->rend_bs,cstart_orig=baij->cstart_bs;
322: PetscInt cend_orig=baij->cend_bs,bs=mat->bs;
324: /* Some Variables required in the macro */
325: Mat A = baij->A;
326: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(A)->data;
327: PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
328: MatScalar *aa=a->a;
330: Mat B = baij->B;
331: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data;
332: PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
333: MatScalar *ba=b->a;
335: PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol;
336: PetscInt low,high,t,ridx,cidx,bs2=a->bs2;
337: MatScalar *ap,*bap;
339: /* for stash */
340: PetscInt n_loc, *in_loc=0;
341: MatScalar *v_loc=0;
345: if(!baij->donotstash){
346: PetscMalloc(n*sizeof(PetscInt),&in_loc);
347: PetscMalloc(n*sizeof(MatScalar),&v_loc);
348: }
350: for (i=0; i<m; i++) {
351: if (im[i] < 0) continue;
352: #if defined(PETSC_USE_BOPT_g)
353: if (im[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->M-1);
354: #endif
355: if (im[i] >= rstart_orig && im[i] < rend_orig) { /* this processor entry */
356: row = im[i] - rstart_orig; /* local row index */
357: for (j=0; j<n; j++) {
358: if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */
359: if (in[j] >= cstart_orig && in[j] < cend_orig){ /* diag entry (A) */
360: col = in[j] - cstart_orig; /* local col index */
361: brow = row/bs; bcol = col/bs;
362: if (brow > bcol) continue; /* ignore lower triangular blocks of A */
363: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
364: MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv);
365: /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
366: } else if (in[j] < 0) continue;
367: #if defined(PETSC_USE_BOPT_g)
368: else if (in[j] >= mat->N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->N-1);}
369: #endif
370: else { /* off-diag entry (B) */
371: if (mat->was_assembled) {
372: if (!baij->colmap) {
373: CreateColmap_MPIBAIJ_Private(mat);
374: }
375: #if defined (PETSC_USE_CTABLE)
376: PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
377: col = col - 1;
378: #else
379: col = baij->colmap[in[j]/bs] - 1;
380: #endif
381: if (col < 0 && !((Mat_SeqSBAIJ*)(baij->A->data))->nonew) {
382: DisAssemble_MPISBAIJ(mat);
383: col = in[j];
384: /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
385: B = baij->B;
386: b = (Mat_SeqBAIJ*)(B)->data;
387: bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
388: ba=b->a;
389: } else col += in[j]%bs;
390: } else col = in[j];
391: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
392: MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv);
393: /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
394: }
395: }
396: } else { /* off processor entry */
397: if (!baij->donotstash) {
398: n_loc = 0;
399: for (j=0; j<n; j++){
400: if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */
401: in_loc[n_loc] = in[j];
402: if (roworiented) {
403: v_loc[n_loc] = v[i*n+j];
404: } else {
405: v_loc[n_loc] = v[j*m+i];
406: }
407: n_loc++;
408: }
409: MatStashValuesRow_Private(&mat->stash,im[i],n_loc,in_loc,v_loc);
410: }
411: }
412: }
414: if(!baij->donotstash){
415: PetscFree(in_loc);
416: PetscFree(v_loc);
417: }
418: return(0);
419: }
423: PetscErrorCode MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const MatScalar v[],InsertMode addv)
424: {
425: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
426: const MatScalar *value;
427: MatScalar *barray=baij->barray;
428: PetscTruth roworiented = baij->roworiented;
429: PetscErrorCode ierr;
430: PetscInt i,j,ii,jj,row,col,rstart=baij->rstart;
431: PetscInt rend=baij->rend,cstart=baij->cstart,stepval;
432: PetscInt cend=baij->cend,bs=mat->bs,bs2=baij->bs2;
435: if(!barray) {
436: PetscMalloc(bs2*sizeof(MatScalar),&barray);
437: baij->barray = barray;
438: }
440: if (roworiented) {
441: stepval = (n-1)*bs;
442: } else {
443: stepval = (m-1)*bs;
444: }
445: for (i=0; i<m; i++) {
446: if (im[i] < 0) continue;
447: #if defined(PETSC_USE_BOPT_g)
448: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
449: #endif
450: if (im[i] >= rstart && im[i] < rend) {
451: row = im[i] - rstart;
452: for (j=0; j<n; j++) {
453: /* If NumCol = 1 then a copy is not required */
454: if ((roworiented) && (n == 1)) {
455: barray = (MatScalar*) v + i*bs2;
456: } else if((!roworiented) && (m == 1)) {
457: barray = (MatScalar*) v + j*bs2;
458: } else { /* Here a copy is required */
459: if (roworiented) {
460: value = v + i*(stepval+bs)*bs + j*bs;
461: } else {
462: value = v + j*(stepval+bs)*bs + i*bs;
463: }
464: for (ii=0; ii<bs; ii++,value+=stepval) {
465: for (jj=0; jj<bs; jj++) {
466: *barray++ = *value++;
467: }
468: }
469: barray -=bs2;
470: }
471:
472: if (in[j] >= cstart && in[j] < cend){
473: col = in[j] - cstart;
474: MatSetValuesBlocked_SeqSBAIJ(baij->A,1,&row,1,&col,barray,addv);
475: }
476: else if (in[j] < 0) continue;
477: #if defined(PETSC_USE_BOPT_g)
478: else if (in[j] >= baij->Nbs) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);}
479: #endif
480: else {
481: if (mat->was_assembled) {
482: if (!baij->colmap) {
483: CreateColmap_MPIBAIJ_Private(mat);
484: }
486: #if defined(PETSC_USE_BOPT_g)
487: #if defined (PETSC_USE_CTABLE)
488: { PetscInt data;
489: PetscTableFind(baij->colmap,in[j]+1,&data);
490: if ((data - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
491: }
492: #else
493: if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
494: #endif
495: #endif
496: #if defined (PETSC_USE_CTABLE)
497: PetscTableFind(baij->colmap,in[j]+1,&col);
498: col = (col - 1)/bs;
499: #else
500: col = (baij->colmap[in[j]] - 1)/bs;
501: #endif
502: if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
503: DisAssemble_MPISBAIJ(mat);
504: col = in[j];
505: }
506: }
507: else col = in[j];
508: MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
509: }
510: }
511: } else {
512: if (!baij->donotstash) {
513: if (roworiented) {
514: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
515: } else {
516: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
517: }
518: }
519: }
520: }
521: return(0);
522: }
524: #define HASH_KEY 0.6180339887
525: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp)))
526: /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
527: /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
530: PetscErrorCode MatSetValues_MPISBAIJ_HT_MatScalar(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const MatScalar v[],InsertMode addv)
531: {
533: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
534: /* return(0); */
535: }
539: PetscErrorCode MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const MatScalar v[],InsertMode addv)
540: {
542: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
543: /* return(0); */
544: }
548: PetscErrorCode MatGetValues_MPISBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
549: {
550: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
552: PetscInt bs=mat->bs,i,j,bsrstart = baij->rstart*bs,bsrend = baij->rend*bs;
553: PetscInt bscstart = baij->cstart*bs,bscend = baij->cend*bs,row,col,data;
556: for (i=0; i<m; i++) {
557: if (idxm[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);
558: if (idxm[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->M-1);
559: if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
560: row = idxm[i] - bsrstart;
561: for (j=0; j<n; j++) {
562: if (idxn[j] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column %D",idxn[j]);
563: if (idxn[j] >= mat->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->N-1);
564: if (idxn[j] >= bscstart && idxn[j] < bscend){
565: col = idxn[j] - bscstart;
566: MatGetValues_SeqSBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
567: } else {
568: if (!baij->colmap) {
569: CreateColmap_MPIBAIJ_Private(mat);
570: }
571: #if defined (PETSC_USE_CTABLE)
572: PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
573: data --;
574: #else
575: data = baij->colmap[idxn[j]/bs]-1;
576: #endif
577: if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
578: else {
579: col = data + idxn[j]%bs;
580: MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
581: }
582: }
583: }
584: } else {
585: SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
586: }
587: }
588: return(0);
589: }
593: PetscErrorCode MatNorm_MPISBAIJ(Mat mat,NormType type,PetscReal *norm)
594: {
595: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
597: PetscReal sum[2],*lnorm2;
600: if (baij->size == 1) {
601: MatNorm(baij->A,type,norm);
602: } else {
603: if (type == NORM_FROBENIUS) {
604: PetscMalloc(2*sizeof(PetscReal),&lnorm2);
605: MatNorm(baij->A,type,lnorm2);
606: *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2++; /* squar power of norm(A) */
607: MatNorm(baij->B,type,lnorm2);
608: *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2--; /* squar power of norm(B) */
609: MPI_Allreduce(lnorm2,&sum,2,MPIU_REAL,MPI_SUM,mat->comm);
610: *norm = sqrt(sum[0] + 2*sum[1]);
611: PetscFree(lnorm2);
612: } else {
613: SETERRQ(PETSC_ERR_SUP,"No support for this norm yet");
614: }
615: }
616: return(0);
617: }
619: /*
620: Creates the hash table, and sets the table
621: This table is created only once.
622: If new entried need to be added to the matrix
623: then the hash table has to be destroyed and
624: recreated.
625: */
628: PetscErrorCode MatCreateHashTable_MPISBAIJ_Private(Mat mat,PetscReal factor)
629: {
631: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
632: /* return(0); */
633: }
637: PetscErrorCode MatAssemblyBegin_MPISBAIJ(Mat mat,MatAssemblyType mode)
638: {
639: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
641: PetscInt nstash,reallocs;
642: InsertMode addv;
645: if (baij->donotstash) {
646: return(0);
647: }
649: /* make sure all processors are either in INSERTMODE or ADDMODE */
650: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);
651: if (addv == (ADD_VALUES|INSERT_VALUES)) {
652: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
653: }
654: mat->insertmode = addv; /* in case this processor had no cache */
656: MatStashScatterBegin_Private(&mat->stash,baij->rowners_bs);
657: MatStashScatterBegin_Private(&mat->bstash,baij->rowners);
658: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
659: PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);
660: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
661: PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
662: return(0);
663: }
667: PetscErrorCode MatAssemblyEnd_MPISBAIJ(Mat mat,MatAssemblyType mode)
668: {
669: Mat_MPISBAIJ *baij=(Mat_MPISBAIJ*)mat->data;
670: Mat_SeqSBAIJ *a=(Mat_SeqSBAIJ*)baij->A->data;
671: Mat_SeqBAIJ *b=(Mat_SeqBAIJ*)baij->B->data;
673: PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2;
674: PetscInt *row,*col,other_disassembled;
675: PetscMPIInt n;
676: PetscTruth r1,r2,r3;
677: MatScalar *val;
678: InsertMode addv = mat->insertmode;
682: if (!baij->donotstash) {
683: while (1) {
684: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
685: if (!flg) break;
687: for (i=0; i<n;) {
688: /* Now identify the consecutive vals belonging to the same row */
689: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
690: if (j < n) ncols = j-i;
691: else ncols = n-i;
692: /* Now assemble all these values with a single function call */
693: MatSetValues_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i,addv);
694: i = j;
695: }
696: }
697: MatStashScatterEnd_Private(&mat->stash);
698: /* Now process the block-stash. Since the values are stashed column-oriented,
699: set the roworiented flag to column oriented, and after MatSetValues()
700: restore the original flags */
701: r1 = baij->roworiented;
702: r2 = a->roworiented;
703: r3 = b->roworiented;
704: baij->roworiented = PETSC_FALSE;
705: a->roworiented = PETSC_FALSE;
706: b->roworiented = PETSC_FALSE;
707: while (1) {
708: MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
709: if (!flg) break;
710:
711: for (i=0; i<n;) {
712: /* Now identify the consecutive vals belonging to the same row */
713: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
714: if (j < n) ncols = j-i;
715: else ncols = n-i;
716: MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
717: i = j;
718: }
719: }
720: MatStashScatterEnd_Private(&mat->bstash);
721: baij->roworiented = r1;
722: a->roworiented = r2;
723: b->roworiented = r3;
724: }
726: MatAssemblyBegin(baij->A,mode);
727: MatAssemblyEnd(baij->A,mode);
729: /* determine if any processor has disassembled, if so we must
730: also disassemble ourselfs, in order that we may reassemble. */
731: /*
732: if nonzero structure of submatrix B cannot change then we know that
733: no processor disassembled thus we can skip this stuff
734: */
735: if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
736: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);
737: if (mat->was_assembled && !other_disassembled) {
738: DisAssemble_MPISBAIJ(mat);
739: }
740: }
742: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
743: MatSetUpMultiply_MPISBAIJ(mat); /* setup Mvctx and sMvctx */
744: }
745: MatAssemblyBegin(baij->B,mode);
746: MatAssemblyEnd(baij->B,mode);
747:
748: #if defined(PETSC_USE_BOPT_g)
749: if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
750: PetscLogInfo(0,"MatAssemblyEnd_MPISBAIJ:Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);
751: baij->ht_total_ct = 0;
752: baij->ht_insert_ct = 0;
753: }
754: #endif
755: if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
756: MatCreateHashTable_MPISBAIJ_Private(mat,baij->ht_fact);
757: mat->ops->setvalues = MatSetValues_MPISBAIJ_HT;
758: mat->ops->setvaluesblocked = MatSetValuesBlocked_MPISBAIJ_HT;
759: }
761: if (baij->rowvalues) {
762: PetscFree(baij->rowvalues);
763: baij->rowvalues = 0;
764: }
766: return(0);
767: }
771: static PetscErrorCode MatView_MPISBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
772: {
773: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
774: PetscErrorCode ierr;
775: PetscInt bs = mat->bs;
776: PetscMPIInt size = baij->size,rank = baij->rank;
777: PetscTruth iascii,isdraw;
778: PetscViewer sviewer;
779: PetscViewerFormat format;
782: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
783: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
784: if (iascii) {
785: PetscViewerGetFormat(viewer,&format);
786: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
787: MatInfo info;
788: MPI_Comm_rank(mat->comm,&rank);
789: MatGetInfo(mat,MAT_LOCAL,&info);
790: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n",
791: rank,mat->m,(PetscInt)info.nz_used*bs,(PetscInt)info.nz_allocated*bs,
792: mat->bs,(PetscInt)info.memory);
793: MatGetInfo(baij->A,MAT_LOCAL,&info);
794: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used*bs);
795: MatGetInfo(baij->B,MAT_LOCAL,&info);
796: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used*bs);
797: PetscViewerFlush(viewer);
798: VecScatterView(baij->Mvctx,viewer);
799: return(0);
800: } else if (format == PETSC_VIEWER_ASCII_INFO) {
801: PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);
802: return(0);
803: }
804: }
806: if (isdraw) {
807: PetscDraw draw;
808: PetscTruth isnull;
809: PetscViewerDrawGetDraw(viewer,0,&draw);
810: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
811: }
813: if (size == 1) {
814: PetscObjectSetName((PetscObject)baij->A,mat->name);
815: MatView(baij->A,viewer);
816: } else {
817: /* assemble the entire matrix onto first processor. */
818: Mat A;
819: Mat_SeqSBAIJ *Aloc;
820: Mat_SeqBAIJ *Bloc;
821: PetscInt M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
822: MatScalar *a;
824: /* Should this be the same type as mat? */
825: if (!rank) {
826: MatCreate(mat->comm,M,N,M,N,&A);
827: } else {
828: MatCreate(mat->comm,0,0,M,N,&A);
829: }
830: MatSetType(A,MATMPISBAIJ);
831: MatMPISBAIJSetPreallocation(A,mat->bs,0,PETSC_NULL,0,PETSC_NULL);
832: PetscLogObjectParent(mat,A);
834: /* copy over the A part */
835: Aloc = (Mat_SeqSBAIJ*)baij->A->data;
836: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
837: PetscMalloc(bs*sizeof(PetscInt),&rvals);
839: for (i=0; i<mbs; i++) {
840: rvals[0] = bs*(baij->rstart + i);
841: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
842: for (j=ai[i]; j<ai[i+1]; j++) {
843: col = (baij->cstart+aj[j])*bs;
844: for (k=0; k<bs; k++) {
845: MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
846: col++; a += bs;
847: }
848: }
849: }
850: /* copy over the B part */
851: Bloc = (Mat_SeqBAIJ*)baij->B->data;
852: ai = Bloc->i; aj = Bloc->j; a = Bloc->a;
853: for (i=0; i<mbs; i++) {
854: rvals[0] = bs*(baij->rstart + i);
855: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
856: for (j=ai[i]; j<ai[i+1]; j++) {
857: col = baij->garray[aj[j]]*bs;
858: for (k=0; k<bs; k++) {
859: MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
860: col++; a += bs;
861: }
862: }
863: }
864: PetscFree(rvals);
865: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
866: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
867: /*
868: Everyone has to call to draw the matrix since the graphics waits are
869: synchronized across all processors that share the PetscDraw object
870: */
871: PetscViewerGetSingleton(viewer,&sviewer);
872: if (!rank) {
873: PetscObjectSetName((PetscObject)((Mat_MPISBAIJ*)(A->data))->A,mat->name);
874: MatView(((Mat_MPISBAIJ*)(A->data))->A,sviewer);
875: }
876: PetscViewerRestoreSingleton(viewer,&sviewer);
877: MatDestroy(A);
878: }
879: return(0);
880: }
884: PetscErrorCode MatView_MPISBAIJ(Mat mat,PetscViewer viewer)
885: {
887: PetscTruth iascii,isdraw,issocket,isbinary;
890: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
891: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
892: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
893: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
894: if (iascii || isdraw || issocket || isbinary) {
895: MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);
896: } else {
897: SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPISBAIJ matrices",((PetscObject)viewer)->type_name);
898: }
899: return(0);
900: }
904: PetscErrorCode MatDestroy_MPISBAIJ(Mat mat)
905: {
906: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
910: #if defined(PETSC_USE_LOG)
911: PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->M,mat->N);
912: #endif
913: MatStashDestroy_Private(&mat->stash);
914: MatStashDestroy_Private(&mat->bstash);
915: PetscFree(baij->rowners);
916: MatDestroy(baij->A);
917: MatDestroy(baij->B);
918: #if defined (PETSC_USE_CTABLE)
919: if (baij->colmap) {PetscTableDelete(baij->colmap);}
920: #else
921: if (baij->colmap) {PetscFree(baij->colmap);}
922: #endif
923: if (baij->garray) {PetscFree(baij->garray);}
924: if (baij->lvec) {VecDestroy(baij->lvec);}
925: if (baij->Mvctx) {VecScatterDestroy(baij->Mvctx);}
926: if (baij->slvec0) {
927: VecDestroy(baij->slvec0);
928: VecDestroy(baij->slvec0b);
929: }
930: if (baij->slvec1) {
931: VecDestroy(baij->slvec1);
932: VecDestroy(baij->slvec1a);
933: VecDestroy(baij->slvec1b);
934: }
935: if (baij->sMvctx) {VecScatterDestroy(baij->sMvctx);}
936: if (baij->rowvalues) {PetscFree(baij->rowvalues);}
937: if (baij->barray) {PetscFree(baij->barray);}
938: if (baij->hd) {PetscFree(baij->hd);}
939: #if defined(PETSC_USE_MAT_SINGLE)
940: if (baij->setvaluescopy) {PetscFree(baij->setvaluescopy);}
941: #endif
942: PetscFree(baij);
944: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
945: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
946: PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
947: PetscObjectComposeFunction((PetscObject)mat,"MatMPISBAIJSetPreallocation_C","",PETSC_NULL);
948: return(0);
949: }
953: PetscErrorCode MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy)
954: {
955: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
957: PetscInt nt,mbs=a->mbs,bs=A->bs;
958: PetscScalar *x,*from,zero=0.0;
959:
961: VecGetLocalSize(xx,&nt);
962: if (nt != A->n) {
963: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
964: }
965: VecGetLocalSize(yy,&nt);
966: if (nt != A->m) {
967: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
968: }
970: /* diagonal part */
971: (*a->A->ops->mult)(a->A,xx,a->slvec1a);
972: VecSet(&zero,a->slvec1b);
974: /* subdiagonal part */
975: (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);
977: /* copy x into the vec slvec0 */
978: VecGetArray(a->slvec0,&from);
979: VecGetArray(xx,&x);
980: PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));
981: VecRestoreArray(a->slvec0,&from);
982:
983: VecScatterBegin(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
984: VecRestoreArray(xx,&x);
985: VecScatterEnd(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
986:
987: /* supperdiagonal part */
988: (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,yy);
989:
990: return(0);
991: }
995: PetscErrorCode MatMult_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy)
996: {
997: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
999: PetscInt nt;
1002: VecGetLocalSize(xx,&nt);
1003: if (nt != A->n) {
1004: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1005: }
1006: VecGetLocalSize(yy,&nt);
1007: if (nt != A->m) {
1008: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1009: }
1011: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1012: /* do diagonal part */
1013: (*a->A->ops->mult)(a->A,xx,yy);
1014: /* do supperdiagonal part */
1015: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1016: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1017: /* do subdiagonal part */
1018: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1019: VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1020: VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1022: return(0);
1023: }
1027: PetscErrorCode MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1028: {
1029: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1031: PetscInt mbs=a->mbs,bs=A->bs;
1032: PetscScalar *x,*from,zero=0.0;
1033:
1035: /*
1036: PetscSynchronizedPrintf(A->comm," MatMultAdd is called ...\n");
1037: PetscSynchronizedFlush(A->comm);
1038: */
1039: /* diagonal part */
1040: (*a->A->ops->multadd)(a->A,xx,yy,a->slvec1a);
1041: VecSet(&zero,a->slvec1b);
1043: /* subdiagonal part */
1044: (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);
1046: /* copy x into the vec slvec0 */
1047: VecGetArray(a->slvec0,&from);
1048: VecGetArray(xx,&x);
1049: PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));
1050: VecRestoreArray(a->slvec0,&from);
1051:
1052: VecScatterBegin(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
1053: VecRestoreArray(xx,&x);
1054: VecScatterEnd(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
1055:
1056: /* supperdiagonal part */
1057: (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,zz);
1058:
1059: return(0);
1060: }
1064: PetscErrorCode MatMultAdd_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy,Vec zz)
1065: {
1066: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1070: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1071: /* do diagonal part */
1072: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1073: /* do supperdiagonal part */
1074: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1075: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1077: /* do subdiagonal part */
1078: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1079: VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1080: VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1082: return(0);
1083: }
1087: PetscErrorCode MatMultTranspose_MPISBAIJ(Mat A,Vec xx,Vec yy)
1088: {
1092: MatMult(A,xx,yy);
1093: return(0);
1094: }
1098: PetscErrorCode MatMultTransposeAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1099: {
1103: MatMultAdd(A,xx,yy,zz);
1104: return(0);
1105: }
1107: /*
1108: This only works correctly for square matrices where the subblock A->A is the
1109: diagonal block
1110: */
1113: PetscErrorCode MatGetDiagonal_MPISBAIJ(Mat A,Vec v)
1114: {
1115: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1119: /* if (a->M != a->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */
1120: MatGetDiagonal(a->A,v);
1121: return(0);
1122: }
1126: PetscErrorCode MatScale_MPISBAIJ(const PetscScalar *aa,Mat A)
1127: {
1128: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1132: MatScale(aa,a->A);
1133: MatScale(aa,a->B);
1134: return(0);
1135: }
1139: PetscErrorCode MatGetRow_MPISBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1140: {
1141: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
1142: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1144: PetscInt bs = matin->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB;
1145: PetscInt nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs;
1146: PetscInt *cmap,*idx_p,cstart = mat->cstart;
1149: if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1150: mat->getrowactive = PETSC_TRUE;
1152: if (!mat->rowvalues && (idx || v)) {
1153: /*
1154: allocate enough space to hold information from the longest row.
1155: */
1156: Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data;
1157: Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data;
1158: PetscInt max = 1,mbs = mat->mbs,tmp;
1159: for (i=0; i<mbs; i++) {
1160: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */
1161: if (max < tmp) { max = tmp; }
1162: }
1163: PetscMalloc(max*bs2*(sizeof(PetscInt)+sizeof(PetscScalar)),&mat->rowvalues);
1164: mat->rowindices = (PetscInt*)(mat->rowvalues + max*bs2);
1165: }
1166:
1167: if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows")
1168: lrow = row - brstart; /* local row index */
1170: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1171: if (!v) {pvA = 0; pvB = 0;}
1172: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1173: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1174: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1175: nztot = nzA + nzB;
1177: cmap = mat->garray;
1178: if (v || idx) {
1179: if (nztot) {
1180: /* Sort by increasing column numbers, assuming A and B already sorted */
1181: PetscInt imark = -1;
1182: if (v) {
1183: *v = v_p = mat->rowvalues;
1184: for (i=0; i<nzB; i++) {
1185: if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1186: else break;
1187: }
1188: imark = i;
1189: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1190: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1191: }
1192: if (idx) {
1193: *idx = idx_p = mat->rowindices;
1194: if (imark > -1) {
1195: for (i=0; i<imark; i++) {
1196: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1197: }
1198: } else {
1199: for (i=0; i<nzB; i++) {
1200: if (cmap[cworkB[i]/bs] < cstart)
1201: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1202: else break;
1203: }
1204: imark = i;
1205: }
1206: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i];
1207: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1208: }
1209: } else {
1210: if (idx) *idx = 0;
1211: if (v) *v = 0;
1212: }
1213: }
1214: *nz = nztot;
1215: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1216: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1217: return(0);
1218: }
1222: PetscErrorCode MatRestoreRow_MPISBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1223: {
1224: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
1227: if (baij->getrowactive == PETSC_FALSE) {
1228: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1229: }
1230: baij->getrowactive = PETSC_FALSE;
1231: return(0);
1232: }
1236: PetscErrorCode MatZeroEntries_MPISBAIJ(Mat A)
1237: {
1238: Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data;
1242: MatZeroEntries(l->A);
1243: MatZeroEntries(l->B);
1244: return(0);
1245: }
1249: PetscErrorCode MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1250: {
1251: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data;
1252: Mat A = a->A,B = a->B;
1254: PetscReal isend[5],irecv[5];
1257: info->block_size = (PetscReal)matin->bs;
1258: MatGetInfo(A,MAT_LOCAL,info);
1259: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1260: isend[3] = info->memory; isend[4] = info->mallocs;
1261: MatGetInfo(B,MAT_LOCAL,info);
1262: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1263: isend[3] += info->memory; isend[4] += info->mallocs;
1264: if (flag == MAT_LOCAL) {
1265: info->nz_used = isend[0];
1266: info->nz_allocated = isend[1];
1267: info->nz_unneeded = isend[2];
1268: info->memory = isend[3];
1269: info->mallocs = isend[4];
1270: } else if (flag == MAT_GLOBAL_MAX) {
1271: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);
1272: info->nz_used = irecv[0];
1273: info->nz_allocated = irecv[1];
1274: info->nz_unneeded = irecv[2];
1275: info->memory = irecv[3];
1276: info->mallocs = irecv[4];
1277: } else if (flag == MAT_GLOBAL_SUM) {
1278: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);
1279: info->nz_used = irecv[0];
1280: info->nz_allocated = irecv[1];
1281: info->nz_unneeded = irecv[2];
1282: info->memory = irecv[3];
1283: info->mallocs = irecv[4];
1284: } else {
1285: SETERRQ1(PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag);
1286: }
1287: info->rows_global = (PetscReal)A->M;
1288: info->columns_global = (PetscReal)A->N;
1289: info->rows_local = (PetscReal)A->m;
1290: info->columns_local = (PetscReal)A->N;
1291: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1292: info->fill_ratio_needed = 0;
1293: info->factor_mallocs = 0;
1294: return(0);
1295: }
1299: PetscErrorCode MatSetOption_MPISBAIJ(Mat A,MatOption op)
1300: {
1301: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1305: switch (op) {
1306: case MAT_NO_NEW_NONZERO_LOCATIONS:
1307: case MAT_YES_NEW_NONZERO_LOCATIONS:
1308: case MAT_COLUMNS_UNSORTED:
1309: case MAT_COLUMNS_SORTED:
1310: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1311: case MAT_KEEP_ZEROED_ROWS:
1312: case MAT_NEW_NONZERO_LOCATION_ERR:
1313: MatSetOption(a->A,op);
1314: MatSetOption(a->B,op);
1315: break;
1316: case MAT_ROW_ORIENTED:
1317: a->roworiented = PETSC_TRUE;
1318: MatSetOption(a->A,op);
1319: MatSetOption(a->B,op);
1320: break;
1321: case MAT_ROWS_SORTED:
1322: case MAT_ROWS_UNSORTED:
1323: case MAT_YES_NEW_DIAGONALS:
1324: PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n");
1325: break;
1326: case MAT_COLUMN_ORIENTED:
1327: a->roworiented = PETSC_FALSE;
1328: MatSetOption(a->A,op);
1329: MatSetOption(a->B,op);
1330: break;
1331: case MAT_IGNORE_OFF_PROC_ENTRIES:
1332: a->donotstash = PETSC_TRUE;
1333: break;
1334: case MAT_NO_NEW_DIAGONALS:
1335: SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1336: case MAT_USE_HASH_TABLE:
1337: a->ht_flag = PETSC_TRUE;
1338: break;
1339: case MAT_NOT_SYMMETRIC:
1340: case MAT_NOT_STRUCTURALLY_SYMMETRIC:
1341: case MAT_HERMITIAN:
1342: SETERRQ(PETSC_ERR_SUP,"Matrix must be symmetric");
1343: case MAT_SYMMETRIC:
1344: case MAT_STRUCTURALLY_SYMMETRIC:
1345: case MAT_NOT_HERMITIAN:
1346: case MAT_SYMMETRY_ETERNAL:
1347: case MAT_NOT_SYMMETRY_ETERNAL:
1348: break;
1349: default:
1350: SETERRQ(PETSC_ERR_SUP,"unknown option");
1351: }
1352: return(0);
1353: }
1357: PetscErrorCode MatTranspose_MPISBAIJ(Mat A,Mat *B)
1358: {
1361: MatDuplicate(A,MAT_COPY_VALUES,B);
1362: return(0);
1363: }
1367: PetscErrorCode MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr)
1368: {
1369: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
1370: Mat a = baij->A,b = baij->B;
1372: PetscInt s1,s2,s3;
1375: if (ll != rr) {
1376: SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n");
1377: }
1378: MatGetLocalSize(mat,&s2,&s3);
1379: if (rr) {
1380: VecGetLocalSize(rr,&s1);
1381: if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1382: /* Overlap communication with computation. */
1383: VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1384: /*} if (ll) { */
1385: VecGetLocalSize(ll,&s1);
1386: if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1387: (*b->ops->diagonalscale)(b,ll,PETSC_NULL);
1388: /* } */
1389: /* scale the diagonal block */
1390: (*a->ops->diagonalscale)(a,ll,rr);
1392: /* if (rr) { */
1393: /* Do a scatter end and then right scale the off-diagonal block */
1394: VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1395: (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);
1396: }
1397:
1398: return(0);
1399: }
1403: PetscErrorCode MatZeroRows_MPISBAIJ(Mat A,IS is,const PetscScalar *diag)
1404: {
1406: SETERRQ(PETSC_ERR_SUP,"No support for this function yet");
1407: }
1411: PetscErrorCode MatPrintHelp_MPISBAIJ(Mat A)
1412: {
1413: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1414: MPI_Comm comm = A->comm;
1415: static PetscTruth called = PETSC_FALSE;
1416: PetscErrorCode ierr;
1419: if (!a->rank) {
1420: MatPrintHelp_SeqSBAIJ(a->A);
1421: }
1422: if (called) {return(0);} else called = PETSC_TRUE;
1423: (*PetscHelpPrintf)(comm," Options for MATMPISBAIJ matrix format (the defaults):\n");
1424: (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");
1425: return(0);
1426: }
1430: PetscErrorCode MatSetUnfactored_MPISBAIJ(Mat A)
1431: {
1432: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1436: MatSetUnfactored(a->A);
1437: return(0);
1438: }
1440: static PetscErrorCode MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat *);
1444: PetscErrorCode MatEqual_MPISBAIJ(Mat A,Mat B,PetscTruth *flag)
1445: {
1446: Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data;
1447: Mat a,b,c,d;
1448: PetscTruth flg;
1452: a = matA->A; b = matA->B;
1453: c = matB->A; d = matB->B;
1455: MatEqual(a,c,&flg);
1456: if (flg == PETSC_TRUE) {
1457: MatEqual(b,d,&flg);
1458: }
1459: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);
1460: return(0);
1461: }
1465: PetscErrorCode MatSetUpPreallocation_MPISBAIJ(Mat A)
1466: {
1470: MatMPISBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1471: return(0);
1472: }
1476: PetscErrorCode MatGetSubMatrices_MPISBAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
1477: {
1479: PetscInt i;
1480: PetscTruth flg;
1483: for (i=0; i<n; i++) {
1484: ISEqual(irow[i],icol[i],&flg);
1485: if (!flg) {
1486: SETERRQ(PETSC_ERR_SUP,"Can only get symmetric submatrix for MPISBAIJ matrices");
1487: }
1488: }
1489: MatGetSubMatrices_MPIBAIJ(A,n,irow,icol,scall,B);
1490: return(0);
1491: }
1492:
1494: /* -------------------------------------------------------------------*/
1495: static struct _MatOps MatOps_Values = {
1496: MatSetValues_MPISBAIJ,
1497: MatGetRow_MPISBAIJ,
1498: MatRestoreRow_MPISBAIJ,
1499: MatMult_MPISBAIJ,
1500: /* 4*/ MatMultAdd_MPISBAIJ,
1501: MatMultTranspose_MPISBAIJ,
1502: MatMultTransposeAdd_MPISBAIJ,
1503: 0,
1504: 0,
1505: 0,
1506: /*10*/ 0,
1507: 0,
1508: 0,
1509: MatRelax_MPISBAIJ,
1510: MatTranspose_MPISBAIJ,
1511: /*15*/ MatGetInfo_MPISBAIJ,
1512: MatEqual_MPISBAIJ,
1513: MatGetDiagonal_MPISBAIJ,
1514: MatDiagonalScale_MPISBAIJ,
1515: MatNorm_MPISBAIJ,
1516: /*20*/ MatAssemblyBegin_MPISBAIJ,
1517: MatAssemblyEnd_MPISBAIJ,
1518: 0,
1519: MatSetOption_MPISBAIJ,
1520: MatZeroEntries_MPISBAIJ,
1521: /*25*/ MatZeroRows_MPISBAIJ,
1522: 0,
1523: 0,
1524: 0,
1525: 0,
1526: /*30*/ MatSetUpPreallocation_MPISBAIJ,
1527: 0,
1528: 0,
1529: 0,
1530: 0,
1531: /*35*/ MatDuplicate_MPISBAIJ,
1532: 0,
1533: 0,
1534: 0,
1535: 0,
1536: /*40*/ 0,
1537: MatGetSubMatrices_MPISBAIJ,
1538: MatIncreaseOverlap_MPISBAIJ,
1539: MatGetValues_MPISBAIJ,
1540: 0,
1541: /*45*/ MatPrintHelp_MPISBAIJ,
1542: MatScale_MPISBAIJ,
1543: 0,
1544: 0,
1545: 0,
1546: /*50*/ 0,
1547: 0,
1548: 0,
1549: 0,
1550: 0,
1551: /*55*/ 0,
1552: 0,
1553: MatSetUnfactored_MPISBAIJ,
1554: 0,
1555: MatSetValuesBlocked_MPISBAIJ,
1556: /*60*/ 0,
1557: 0,
1558: 0,
1559: MatGetPetscMaps_Petsc,
1560: 0,
1561: /*65*/ 0,
1562: 0,
1563: 0,
1564: 0,
1565: 0,
1566: /*70*/ MatGetRowMax_MPISBAIJ,
1567: 0,
1568: 0,
1569: 0,
1570: 0,
1571: /*75*/ 0,
1572: 0,
1573: 0,
1574: 0,
1575: 0,
1576: /*80*/ 0,
1577: 0,
1578: 0,
1579: 0,
1580: MatLoad_MPISBAIJ,
1581: /*85*/ 0,
1582: 0,
1583: 0,
1584: 0,
1585: 0,
1586: /*90*/ 0,
1587: 0,
1588: 0,
1589: 0,
1590: 0,
1591: /*95*/ 0,
1592: 0,
1593: 0,
1594: 0};
1600: PetscErrorCode MatGetDiagonalBlock_MPISBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a)
1601: {
1603: *a = ((Mat_MPISBAIJ *)A->data)->A;
1604: *iscopy = PETSC_FALSE;
1605: return(0);
1606: }
1612: PetscErrorCode MatMPISBAIJSetPreallocation_MPISBAIJ(Mat B,PetscInt bs,PetscInt d_nz,PetscInt *d_nnz,PetscInt o_nz,PetscInt *o_nnz)
1613: {
1614: Mat_MPISBAIJ *b;
1616: PetscInt i,mbs,Mbs;
1619: PetscOptionsGetInt(B->prefix,"-mat_block_size",&bs,PETSC_NULL);
1621: if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive");
1622: if (d_nz == PETSC_DECIDE || d_nz == PETSC_DEFAULT) d_nz = 3;
1623: if (o_nz == PETSC_DECIDE || o_nz == PETSC_DEFAULT) o_nz = 1;
1624: if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
1625: if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);
1626: if (d_nnz) {
1627: for (i=0; i<B->m/bs; i++) {
1628: 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]);
1629: }
1630: }
1631: if (o_nnz) {
1632: for (i=0; i<B->m/bs; i++) {
1633: 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]);
1634: }
1635: }
1636: B->preallocated = PETSC_TRUE;
1637: PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);
1638: PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);
1639: PetscMapCreateMPI(B->comm,B->m,B->M,&B->rmap);
1640: PetscMapCreateMPI(B->comm,B->m,B->M,&B->cmap);
1642: b = (Mat_MPISBAIJ*)B->data;
1643: mbs = B->m/bs;
1644: Mbs = B->M/bs;
1645: if (mbs*bs != B->m) {
1646: SETERRQ2(PETSC_ERR_ARG_SIZ,"No of local rows %D must be divisible by blocksize %D",B->m,bs);
1647: }
1649: B->bs = bs;
1650: b->bs2 = bs*bs;
1651: b->mbs = mbs;
1652: b->nbs = mbs;
1653: b->Mbs = Mbs;
1654: b->Nbs = Mbs;
1656: MPI_Allgather(&b->mbs,1,MPIU_INT,b->rowners+1,1,MPIU_INT,B->comm);
1657: b->rowners[0] = 0;
1658: for (i=2; i<=b->size; i++) {
1659: b->rowners[i] += b->rowners[i-1];
1660: }
1661: b->rstart = b->rowners[b->rank];
1662: b->rend = b->rowners[b->rank+1];
1663: b->cstart = b->rstart;
1664: b->cend = b->rend;
1665: for (i=0; i<=b->size; i++) {
1666: b->rowners_bs[i] = b->rowners[i]*bs;
1667: }
1668: b->rstart_bs = b-> rstart*bs;
1669: b->rend_bs = b->rend*bs;
1670:
1671: b->cstart_bs = b->cstart*bs;
1672: b->cend_bs = b->cend*bs;
1673:
1674: MatCreate(PETSC_COMM_SELF,B->m,B->m,B->m,B->m,&b->A);
1675: MatSetType(b->A,MATSEQSBAIJ);
1676: MatSeqSBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);
1677: PetscLogObjectParent(B,b->A);
1679: MatCreate(PETSC_COMM_SELF,B->m,B->M,B->m,B->M,&b->B);
1680: MatSetType(b->B,MATSEQBAIJ);
1681: MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);
1682: PetscLogObjectParent(B,b->B);
1684: /* build cache for off array entries formed */
1685: MatStashCreate_Private(B->comm,bs,&B->bstash);
1687: return(0);
1688: }
1691: /*MC
1692: MATMPISBAIJ - MATMPISBAIJ = "mpisbaij" - A matrix type to be used for distributed symmetric sparse block matrices,
1693: based on block compressed sparse row format. Only the upper triangular portion of the matrix is stored.
1695: Options Database Keys:
1696: . -mat_type mpisbaij - sets the matrix type to "mpisbaij" during a call to MatSetFromOptions()
1698: Level: beginner
1700: .seealso: MatCreateMPISBAIJ
1701: M*/
1706: PetscErrorCode MatCreate_MPISBAIJ(Mat B)
1707: {
1708: Mat_MPISBAIJ *b;
1710: PetscTruth flg;
1714: PetscNew(Mat_MPISBAIJ,&b);
1715: B->data = (void*)b;
1716: PetscMemzero(b,sizeof(Mat_MPISBAIJ));
1717: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
1719: B->ops->destroy = MatDestroy_MPISBAIJ;
1720: B->ops->view = MatView_MPISBAIJ;
1721: B->mapping = 0;
1722: B->factor = 0;
1723: B->assembled = PETSC_FALSE;
1725: B->insertmode = NOT_SET_VALUES;
1726: MPI_Comm_rank(B->comm,&b->rank);
1727: MPI_Comm_size(B->comm,&b->size);
1729: /* build local table of row and column ownerships */
1730: PetscMalloc(3*(b->size+2)*sizeof(PetscInt),&b->rowners);
1731: b->cowners = b->rowners + b->size + 2;
1732: b->rowners_bs = b->cowners + b->size + 2;
1733: PetscLogObjectMemory(B,3*(b->size+2)*sizeof(PetscInt)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ));
1735: /* build cache for off array entries formed */
1736: MatStashCreate_Private(B->comm,1,&B->stash);
1737: b->donotstash = PETSC_FALSE;
1738: b->colmap = PETSC_NULL;
1739: b->garray = PETSC_NULL;
1740: b->roworiented = PETSC_TRUE;
1742: #if defined(PETSC_USE_MAT_SINGLE)
1743: /* stuff for MatSetValues_XXX in single precision */
1744: b->setvalueslen = 0;
1745: b->setvaluescopy = PETSC_NULL;
1746: #endif
1748: /* stuff used in block assembly */
1749: b->barray = 0;
1751: /* stuff used for matrix vector multiply */
1752: b->lvec = 0;
1753: b->Mvctx = 0;
1754: b->slvec0 = 0;
1755: b->slvec0b = 0;
1756: b->slvec1 = 0;
1757: b->slvec1a = 0;
1758: b->slvec1b = 0;
1759: b->sMvctx = 0;
1761: /* stuff for MatGetRow() */
1762: b->rowindices = 0;
1763: b->rowvalues = 0;
1764: b->getrowactive = PETSC_FALSE;
1766: /* hash table stuff */
1767: b->ht = 0;
1768: b->hd = 0;
1769: b->ht_size = 0;
1770: b->ht_flag = PETSC_FALSE;
1771: b->ht_fact = 0;
1772: b->ht_total_ct = 0;
1773: b->ht_insert_ct = 0;
1775: PetscOptionsHasName(B->prefix,"-mat_use_hash_table",&flg);
1776: if (flg) {
1777: PetscReal fact = 1.39;
1778: MatSetOption(B,MAT_USE_HASH_TABLE);
1779: PetscOptionsGetReal(B->prefix,"-mat_use_hash_table",&fact,PETSC_NULL);
1780: if (fact <= 1.0) fact = 1.39;
1781: MatMPIBAIJSetHashTableFactor(B,fact);
1782: PetscLogInfo(0,"MatCreateMPISBAIJ:Hash table Factor used %5.2f\n",fact);
1783: }
1784: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
1785: "MatStoreValues_MPISBAIJ",
1786: MatStoreValues_MPISBAIJ);
1787: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
1788: "MatRetrieveValues_MPISBAIJ",
1789: MatRetrieveValues_MPISBAIJ);
1790: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
1791: "MatGetDiagonalBlock_MPISBAIJ",
1792: MatGetDiagonalBlock_MPISBAIJ);
1793: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPISBAIJSetPreallocation_C",
1794: "MatMPISBAIJSetPreallocation_MPISBAIJ",
1795: MatMPISBAIJSetPreallocation_MPISBAIJ);
1796: B->symmetric = PETSC_TRUE;
1797: B->structurally_symmetric = PETSC_TRUE;
1798: B->symmetric_set = PETSC_TRUE;
1799: B->structurally_symmetric_set = PETSC_TRUE;
1800: return(0);
1801: }
1804: /*MC
1805: MATSBAIJ - MATSBAIJ = "sbaij" - A matrix type to be used for symmetric block sparse matrices.
1807: This matrix type is identical to MATSEQSBAIJ when constructed with a single process communicator,
1808: and MATMPISBAIJ otherwise.
1810: Options Database Keys:
1811: . -mat_type sbaij - sets the matrix type to "sbaij" during a call to MatSetFromOptions()
1813: Level: beginner
1815: .seealso: MatCreateMPISBAIJ,MATSEQSBAIJ,MATMPISBAIJ
1816: M*/
1821: PetscErrorCode MatCreate_SBAIJ(Mat A)
1822: {
1824: PetscMPIInt size;
1827: PetscObjectChangeTypeName((PetscObject)A,MATSBAIJ);
1828: MPI_Comm_size(A->comm,&size);
1829: if (size == 1) {
1830: MatSetType(A,MATSEQSBAIJ);
1831: } else {
1832: MatSetType(A,MATMPISBAIJ);
1833: }
1834: return(0);
1835: }
1840: /*@C
1841: MatMPISBAIJSetPreallocation - For good matrix assembly performance
1842: the user should preallocate the matrix storage by setting the parameters
1843: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
1844: performance can be increased by more than a factor of 50.
1846: Collective on Mat
1848: Input Parameters:
1849: + A - the matrix
1850: . bs - size of blockk
1851: . d_nz - number of block nonzeros per block row in diagonal portion of local
1852: submatrix (same for all local rows)
1853: . d_nnz - array containing the number of block nonzeros in the various block rows
1854: in the upper triangular and diagonal part of the in diagonal portion of the local
1855: (possibly different for each block row) or PETSC_NULL. You must leave room
1856: for the diagonal entry even if it is zero.
1857: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
1858: submatrix (same for all local rows).
1859: - o_nnz - array containing the number of nonzeros in the various block rows of the
1860: off-diagonal portion of the local submatrix (possibly different for
1861: each block row) or PETSC_NULL.
1864: Options Database Keys:
1865: . -mat_no_unroll - uses code that does not unroll the loops in the
1866: block calculations (much slower)
1867: . -mat_block_size - size of the blocks to use
1869: Notes:
1871: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
1872: than it must be used on all processors that share the object for that argument.
1874: If the *_nnz parameter is given then the *_nz parameter is ignored
1876: Storage Information:
1877: For a square global matrix we define each processor's diagonal portion
1878: to be its local rows and the corresponding columns (a square submatrix);
1879: each processor's off-diagonal portion encompasses the remainder of the
1880: local matrix (a rectangular submatrix).
1882: The user can specify preallocated storage for the diagonal part of
1883: the local submatrix with either d_nz or d_nnz (not both). Set
1884: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
1885: memory allocation. Likewise, specify preallocated storage for the
1886: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
1888: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
1889: the figure below we depict these three local rows and all columns (0-11).
1891: .vb
1892: 0 1 2 3 4 5 6 7 8 9 10 11
1893: -------------------
1894: row 3 | o o o d d d o o o o o o
1895: row 4 | o o o d d d o o o o o o
1896: row 5 | o o o d d d o o o o o o
1897: -------------------
1898: .ve
1899:
1900: Thus, any entries in the d locations are stored in the d (diagonal)
1901: submatrix, and any entries in the o locations are stored in the
1902: o (off-diagonal) submatrix. Note that the d matrix is stored in
1903: MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format.
1905: Now d_nz should indicate the number of block nonzeros per row in the upper triangular
1906: plus the diagonal part of the d matrix,
1907: and o_nz should indicate the number of block nonzeros per row in the o matrix.
1908: In general, for PDE problems in which most nonzeros are near the diagonal,
1909: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
1910: or you will get TERRIBLE performance; see the users' manual chapter on
1911: matrices.
1913: Level: intermediate
1915: .keywords: matrix, block, aij, compressed row, sparse, parallel
1917: .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
1918: @*/
1919: PetscErrorCode MatMPISBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
1920: {
1921: PetscErrorCode ierr,(*f)(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);
1924: PetscObjectQueryFunction((PetscObject)B,"MatMPISBAIJSetPreallocation_C",(void (**)(void))&f);
1925: if (f) {
1926: (*f)(B,bs,d_nz,d_nnz,o_nz,o_nnz);
1927: }
1928: return(0);
1929: }
1933: /*@C
1934: MatCreateMPISBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format
1935: (block compressed row). For good matrix assembly performance
1936: the user should preallocate the matrix storage by setting the parameters
1937: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
1938: performance can be increased by more than a factor of 50.
1940: Collective on MPI_Comm
1942: Input Parameters:
1943: + comm - MPI communicator
1944: . bs - size of blockk
1945: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
1946: This value should be the same as the local size used in creating the
1947: y vector for the matrix-vector product y = Ax.
1948: . n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
1949: This value should be the same as the local size used in creating the
1950: x vector for the matrix-vector product y = Ax.
1951: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
1952: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
1953: . d_nz - number of block nonzeros per block row in diagonal portion of local
1954: submatrix (same for all local rows)
1955: . d_nnz - array containing the number of block nonzeros in the various block rows
1956: in the upper triangular portion of the in diagonal portion of the local
1957: (possibly different for each block block row) or PETSC_NULL.
1958: You must leave room for the diagonal entry even if it is zero.
1959: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
1960: submatrix (same for all local rows).
1961: - o_nnz - array containing the number of nonzeros in the various block rows of the
1962: off-diagonal portion of the local submatrix (possibly different for
1963: each block row) or PETSC_NULL.
1965: Output Parameter:
1966: . A - the matrix
1968: Options Database Keys:
1969: . -mat_no_unroll - uses code that does not unroll the loops in the
1970: block calculations (much slower)
1971: . -mat_block_size - size of the blocks to use
1972: . -mat_mpi - use the parallel matrix data structures even on one processor
1973: (defaults to using SeqBAIJ format on one processor)
1975: Notes:
1976: The user MUST specify either the local or global matrix dimensions
1977: (possibly both).
1979: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
1980: than it must be used on all processors that share the object for that argument.
1982: If the *_nnz parameter is given then the *_nz parameter is ignored
1984: Storage Information:
1985: For a square global matrix we define each processor's diagonal portion
1986: to be its local rows and the corresponding columns (a square submatrix);
1987: each processor's off-diagonal portion encompasses the remainder of the
1988: local matrix (a rectangular submatrix).
1990: The user can specify preallocated storage for the diagonal part of
1991: the local submatrix with either d_nz or d_nnz (not both). Set
1992: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
1993: memory allocation. Likewise, specify preallocated storage for the
1994: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
1996: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
1997: the figure below we depict these three local rows and all columns (0-11).
1999: .vb
2000: 0 1 2 3 4 5 6 7 8 9 10 11
2001: -------------------
2002: row 3 | o o o d d d o o o o o o
2003: row 4 | o o o d d d o o o o o o
2004: row 5 | o o o d d d o o o o o o
2005: -------------------
2006: .ve
2007:
2008: Thus, any entries in the d locations are stored in the d (diagonal)
2009: submatrix, and any entries in the o locations are stored in the
2010: o (off-diagonal) submatrix. Note that the d matrix is stored in
2011: MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format.
2013: Now d_nz should indicate the number of block nonzeros per row in the upper triangular
2014: plus the diagonal part of the d matrix,
2015: and o_nz should indicate the number of block nonzeros per row in the o matrix.
2016: In general, for PDE problems in which most nonzeros are near the diagonal,
2017: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
2018: or you will get TERRIBLE performance; see the users' manual chapter on
2019: matrices.
2021: Level: intermediate
2023: .keywords: matrix, block, aij, compressed row, sparse, parallel
2025: .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
2026: @*/
2028: PetscErrorCode MatCreateMPISBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
2029: {
2031: PetscMPIInt size;
2034: MatCreate(comm,m,n,M,N,A);
2035: MPI_Comm_size(comm,&size);
2036: if (size > 1) {
2037: MatSetType(*A,MATMPISBAIJ);
2038: MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
2039: } else {
2040: MatSetType(*A,MATSEQSBAIJ);
2041: MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
2042: }
2043: return(0);
2044: }
2049: static PetscErrorCode MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2050: {
2051: Mat mat;
2052: Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data;
2054: PetscInt len=0,nt,bs=matin->bs,mbs=oldmat->mbs;
2055: PetscScalar *array;
2058: *newmat = 0;
2059: MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);
2060: MatSetType(mat,matin->type_name);
2061: PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
2062:
2063: mat->factor = matin->factor;
2064: mat->preallocated = PETSC_TRUE;
2065: mat->assembled = PETSC_TRUE;
2066: mat->insertmode = NOT_SET_VALUES;
2068: a = (Mat_MPISBAIJ*)mat->data;
2069: mat->bs = matin->bs;
2070: a->bs2 = oldmat->bs2;
2071: a->mbs = oldmat->mbs;
2072: a->nbs = oldmat->nbs;
2073: a->Mbs = oldmat->Mbs;
2074: a->Nbs = oldmat->Nbs;
2075:
2076: a->rstart = oldmat->rstart;
2077: a->rend = oldmat->rend;
2078: a->cstart = oldmat->cstart;
2079: a->cend = oldmat->cend;
2080: a->size = oldmat->size;
2081: a->rank = oldmat->rank;
2082: a->donotstash = oldmat->donotstash;
2083: a->roworiented = oldmat->roworiented;
2084: a->rowindices = 0;
2085: a->rowvalues = 0;
2086: a->getrowactive = PETSC_FALSE;
2087: a->barray = 0;
2088: a->rstart_bs = oldmat->rstart_bs;
2089: a->rend_bs = oldmat->rend_bs;
2090: a->cstart_bs = oldmat->cstart_bs;
2091: a->cend_bs = oldmat->cend_bs;
2093: /* hash table stuff */
2094: a->ht = 0;
2095: a->hd = 0;
2096: a->ht_size = 0;
2097: a->ht_flag = oldmat->ht_flag;
2098: a->ht_fact = oldmat->ht_fact;
2099: a->ht_total_ct = 0;
2100: a->ht_insert_ct = 0;
2101:
2102: PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(PetscInt));
2103: MatStashCreate_Private(matin->comm,1,&mat->stash);
2104: MatStashCreate_Private(matin->comm,matin->bs,&mat->bstash);
2105: if (oldmat->colmap) {
2106: #if defined (PETSC_USE_CTABLE)
2107: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2108: #else
2109: PetscMalloc((a->Nbs)*sizeof(PetscInt),&a->colmap);
2110: PetscLogObjectMemory(mat,(a->Nbs)*sizeof(PetscInt));
2111: PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));
2112: #endif
2113: } else a->colmap = 0;
2115: if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
2116: PetscMalloc(len*sizeof(PetscInt),&a->garray);
2117: PetscLogObjectMemory(mat,len*sizeof(PetscInt));
2118: PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));
2119: } else a->garray = 0;
2120:
2121: VecDuplicate(oldmat->lvec,&a->lvec);
2122: PetscLogObjectParent(mat,a->lvec);
2123: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2124: PetscLogObjectParent(mat,a->Mvctx);
2126: VecDuplicate(oldmat->slvec0,&a->slvec0);
2127: PetscLogObjectParent(mat,a->slvec0);
2128: VecDuplicate(oldmat->slvec1,&a->slvec1);
2129: PetscLogObjectParent(mat,a->slvec1);
2131: VecGetLocalSize(a->slvec1,&nt);
2132: VecGetArray(a->slvec1,&array);
2133: VecCreateSeqWithArray(PETSC_COMM_SELF,bs*mbs,array,&a->slvec1a);
2134: VecCreateSeqWithArray(PETSC_COMM_SELF,nt-bs*mbs,array+bs*mbs,&a->slvec1b);
2135: VecRestoreArray(a->slvec1,&array);
2136: VecGetArray(a->slvec0,&array);
2137: VecCreateSeqWithArray(PETSC_COMM_SELF,nt-bs*mbs,array+bs*mbs,&a->slvec0b);
2138: VecRestoreArray(a->slvec0,&array);
2139: PetscLogObjectParent(mat,a->slvec0);
2140: PetscLogObjectParent(mat,a->slvec1);
2141: PetscLogObjectParent(mat,a->slvec0b);
2142: PetscLogObjectParent(mat,a->slvec1a);
2143: PetscLogObjectParent(mat,a->slvec1b);
2145: /* VecScatterCopy(oldmat->sMvctx,&a->sMvctx); - not written yet, replaced by the lazy trick: */
2146: PetscObjectReference((PetscObject)oldmat->sMvctx);
2147: a->sMvctx = oldmat->sMvctx;
2148: PetscLogObjectParent(mat,a->sMvctx);
2150: MatDuplicate(oldmat->A,cpvalues,&a->A);
2151: PetscLogObjectParent(mat,a->A);
2152: MatDuplicate(oldmat->B,cpvalues,&a->B);
2153: PetscLogObjectParent(mat,a->B);
2154: PetscFListDuplicate(mat->qlist,&matin->qlist);
2155: *newmat = mat;
2156: return(0);
2157: }
2159: #include petscsys.h
2163: PetscErrorCode MatLoad_MPISBAIJ(PetscViewer viewer,const MatType type,Mat *newmat)
2164: {
2165: Mat A;
2167: PetscInt i,nz,j,rstart,rend;
2168: PetscScalar *vals,*buf;
2169: MPI_Comm comm = ((PetscObject)viewer)->comm;
2170: MPI_Status status;
2171: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag,*sndcounts = 0,*browners,maxnz,*rowners;
2172: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
2173: PetscInt *locrowlens,*procsnz = 0,jj,*mycols,*ibuf;
2174: PetscInt bs=1,Mbs,mbs,extra_rows;
2175: PetscInt *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount;
2176: PetscInt dcount,kmax,k,nzcount,tmp;
2177: int fd;
2178:
2180: PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);
2182: MPI_Comm_size(comm,&size);
2183: MPI_Comm_rank(comm,&rank);
2184: if (!rank) {
2185: PetscViewerBinaryGetDescriptor(viewer,&fd);
2186: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
2187: if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2188: if (header[3] < 0) {
2189: SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ");
2190: }
2191: }
2193: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2194: M = header[1]; N = header[2];
2196: if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices");
2198: /*
2199: This code adds extra rows to make sure the number of rows is
2200: divisible by the blocksize
2201: */
2202: Mbs = M/bs;
2203: extra_rows = bs - M + bs*(Mbs);
2204: if (extra_rows == bs) extra_rows = 0;
2205: else Mbs++;
2206: if (extra_rows &&!rank) {
2207: PetscLogInfo(0,"MatLoad_MPISBAIJ:Padding loaded matrix to match blocksize\n");
2208: }
2210: /* determine ownership of all rows */
2211: mbs = Mbs/size + ((Mbs % size) > rank);
2212: m = mbs*bs;
2213: PetscMalloc(2*(size+2)*sizeof(PetscMPIInt),&rowners);
2214: browners = rowners + size + 1;
2215: MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);
2216: rowners[0] = 0;
2217: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
2218: for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
2219: rstart = rowners[rank];
2220: rend = rowners[rank+1];
2221:
2222: /* distribute row lengths to all processors */
2223: PetscMalloc((rend-rstart)*bs*sizeof(PetscInt),&locrowlens);
2224: if (!rank) {
2225: PetscMalloc((M+extra_rows)*sizeof(PetscInt),&rowlengths);
2226: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
2227: for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1;
2228: PetscMalloc(size*sizeof(PetscMPIInt),&sndcounts);
2229: for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i];
2230: MPI_Scatterv(rowlengths,sndcounts,browners,MPIU_INT,locrowlens,(rend-rstart)*bs,MPIU_INT,0,comm);
2231: PetscFree(sndcounts);
2232: } else {
2233: MPI_Scatterv(0,0,0,MPIU_INT,locrowlens,(rend-rstart)*bs,MPIU_INT,0,comm);
2234: }
2235:
2236: if (!rank) { /* procs[0] */
2237: /* calculate the number of nonzeros on each processor */
2238: PetscMalloc(size*sizeof(PetscInt),&procsnz);
2239: PetscMemzero(procsnz,size*sizeof(PetscInt));
2240: for (i=0; i<size; i++) {
2241: for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) {
2242: procsnz[i] += rowlengths[j];
2243: }
2244: }
2245: PetscFree(rowlengths);
2246:
2247: /* determine max buffer needed and allocate it */
2248: maxnz = 0;
2249: for (i=0; i<size; i++) {
2250: maxnz = PetscMax(maxnz,procsnz[i]);
2251: }
2252: PetscMalloc(maxnz*sizeof(PetscInt),&cols);
2254: /* read in my part of the matrix column indices */
2255: nz = procsnz[0];
2256: PetscMalloc(nz*sizeof(PetscInt),&ibuf);
2257: mycols = ibuf;
2258: if (size == 1) nz -= extra_rows;
2259: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
2260: if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; }
2262: /* read in every ones (except the last) and ship off */
2263: for (i=1; i<size-1; i++) {
2264: nz = procsnz[i];
2265: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2266: MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
2267: }
2268: /* read in the stuff for the last proc */
2269: if (size != 1) {
2270: nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */
2271: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2272: for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
2273: MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);
2274: }
2275: PetscFree(cols);
2276: } else { /* procs[i], i>0 */
2277: /* determine buffer space needed for message */
2278: nz = 0;
2279: for (i=0; i<m; i++) {
2280: nz += locrowlens[i];
2281: }
2282: PetscMalloc(nz*sizeof(PetscInt),&ibuf);
2283: mycols = ibuf;
2284: /* receive message of column indices*/
2285: MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
2286: MPI_Get_count(&status,MPIU_INT,&maxnz);
2287: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2288: }
2290: /* loop over local rows, determining number of off diagonal entries */
2291: PetscMalloc(2*(rend-rstart+1)*sizeof(PetscInt),&dlens);
2292: odlens = dlens + (rend-rstart);
2293: PetscMalloc(3*Mbs*sizeof(PetscInt),&mask);
2294: PetscMemzero(mask,3*Mbs*sizeof(PetscInt));
2295: masked1 = mask + Mbs;
2296: masked2 = masked1 + Mbs;
2297: rowcount = 0; nzcount = 0;
2298: for (i=0; i<mbs; i++) {
2299: dcount = 0;
2300: odcount = 0;
2301: for (j=0; j<bs; j++) {
2302: kmax = locrowlens[rowcount];
2303: for (k=0; k<kmax; k++) {
2304: tmp = mycols[nzcount++]/bs; /* block col. index */
2305: if (!mask[tmp]) {
2306: mask[tmp] = 1;
2307: if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */
2308: else masked1[dcount++] = tmp; /* entry in diag portion */
2309: }
2310: }
2311: rowcount++;
2312: }
2313:
2314: dlens[i] = dcount; /* d_nzz[i] */
2315: odlens[i] = odcount; /* o_nzz[i] */
2317: /* zero out the mask elements we set */
2318: for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
2319: for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
2320: }
2321:
2322: /* create our matrix */
2323: MatCreate(comm,m,m,PETSC_DETERMINE,PETSC_DETERMINE,&A);
2324: MatSetType(A,type);
2325: MatMPISBAIJSetPreallocation(A,bs,0,dlens,0,odlens);
2326: MatSetOption(A,MAT_COLUMNS_SORTED);
2327:
2328: if (!rank) {
2329: PetscMalloc(maxnz*sizeof(PetscScalar),&buf);
2330: /* read in my part of the matrix numerical values */
2331: nz = procsnz[0];
2332: vals = buf;
2333: mycols = ibuf;
2334: if (size == 1) nz -= extra_rows;
2335: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2336: if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; }
2338: /* insert into matrix */
2339: jj = rstart*bs;
2340: for (i=0; i<m; i++) {
2341: MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2342: mycols += locrowlens[i];
2343: vals += locrowlens[i];
2344: jj++;
2345: }
2347: /* read in other processors (except the last one) and ship out */
2348: for (i=1; i<size-1; i++) {
2349: nz = procsnz[i];
2350: vals = buf;
2351: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2352: MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);
2353: }
2354: /* the last proc */
2355: if (size != 1){
2356: nz = procsnz[i] - extra_rows;
2357: vals = buf;
2358: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2359: for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
2360: MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);
2361: }
2362: PetscFree(procsnz);
2364: } else {
2365: /* receive numeric values */
2366: PetscMalloc(nz*sizeof(PetscScalar),&buf);
2368: /* receive message of values*/
2369: vals = buf;
2370: mycols = ibuf;
2371: MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);
2372: MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2373: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2375: /* insert into matrix */
2376: jj = rstart*bs;
2377: for (i=0; i<m; i++) {
2378: MatSetValues_MPISBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2379: mycols += locrowlens[i];
2380: vals += locrowlens[i];
2381: jj++;
2382: }
2383: }
2385: PetscFree(locrowlens);
2386: PetscFree(buf);
2387: PetscFree(ibuf);
2388: PetscFree(rowners);
2389: PetscFree(dlens);
2390: PetscFree(mask);
2391: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2392: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2393: *newmat = A;
2394: return(0);
2395: }
2399: /*@
2400: MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.
2402: Input Parameters:
2403: . mat - the matrix
2404: . fact - factor
2406: Collective on Mat
2408: Level: advanced
2410: Notes:
2411: This can also be set by the command line option: -mat_use_hash_table fact
2413: .keywords: matrix, hashtable, factor, HT
2415: .seealso: MatSetOption()
2416: @*/
2417: PetscErrorCode MatMPISBAIJSetHashTableFactor(Mat mat,PetscReal fact)
2418: {
2420: SETERRQ(PETSC_ERR_SUP,"Function not yet written for SBAIJ format");
2421: /* return(0); */
2422: }
2426: PetscErrorCode MatGetRowMax_MPISBAIJ(Mat A,Vec v)
2427: {
2428: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
2429: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data;
2430: PetscReal atmp;
2431: PetscReal *work,*svalues,*rvalues;
2433: PetscInt i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol;
2434: PetscMPIInt rank,size;
2435: PetscInt *rowners_bs,dest,count,source;
2436: PetscScalar *va;
2437: MatScalar *ba;
2438: MPI_Status stat;
2441: MatGetRowMax(a->A,v);
2442: VecGetArray(v,&va);
2444: MPI_Comm_size(A->comm,&size);
2445: MPI_Comm_rank(A->comm,&rank);
2447: bs = A->bs;
2448: mbs = a->mbs;
2449: Mbs = a->Mbs;
2450: ba = b->a;
2451: bi = b->i;
2452: bj = b->j;
2454: /* find ownerships */
2455: rowners_bs = a->rowners_bs;
2457: /* each proc creates an array to be distributed */
2458: PetscMalloc(bs*Mbs*sizeof(PetscReal),&work);
2459: PetscMemzero(work,bs*Mbs*sizeof(PetscReal));
2461: /* row_max for B */
2462: if (rank != size-1){
2463: for (i=0; i<mbs; i++) {
2464: ncols = bi[1] - bi[0]; bi++;
2465: brow = bs*i;
2466: for (j=0; j<ncols; j++){
2467: bcol = bs*(*bj);
2468: for (kcol=0; kcol<bs; kcol++){
2469: col = bcol + kcol; /* local col index */
2470: col += rowners_bs[rank+1]; /* global col index */
2471: for (krow=0; krow<bs; krow++){
2472: atmp = PetscAbsScalar(*ba); ba++;
2473: row = brow + krow; /* local row index */
2474: if (PetscRealPart(va[row]) < atmp) va[row] = atmp;
2475: if (work[col] < atmp) work[col] = atmp;
2476: }
2477: }
2478: bj++;
2479: }
2480: }
2482: /* send values to its owners */
2483: for (dest=rank+1; dest<size; dest++){
2484: svalues = work + rowners_bs[dest];
2485: count = rowners_bs[dest+1]-rowners_bs[dest];
2486: MPI_Send(svalues,count,MPIU_REAL,dest,rank,A->comm);
2487: }
2488: }
2489:
2490: /* receive values */
2491: if (rank){
2492: rvalues = work;
2493: count = rowners_bs[rank+1]-rowners_bs[rank];
2494: for (source=0; source<rank; source++){
2495: MPI_Recv(rvalues,count,MPIU_REAL,MPI_ANY_SOURCE,MPI_ANY_TAG,A->comm,&stat);
2496: /* process values */
2497: for (i=0; i<count; i++){
2498: if (PetscRealPart(va[i]) < rvalues[i]) va[i] = rvalues[i];
2499: }
2500: }
2501: }
2503: VecRestoreArray(v,&va);
2504: PetscFree(work);
2505: return(0);
2506: }
2510: PetscErrorCode MatRelax_MPISBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
2511: {
2512: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
2514: PetscInt mbs=mat->mbs,bs=matin->bs;
2515: PetscScalar mone=-1.0,*x,*b,*ptr,zero=0.0;
2516: Vec bb1;
2517:
2519: if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
2520: if (bs > 1)
2521: SETERRQ(PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented");
2523: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
2524: if ( flag & SOR_ZERO_INITIAL_GUESS ) {
2525: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
2526: its--;
2527: }
2529: VecDuplicate(bb,&bb1);
2530: while (its--){
2531:
2532: /* lower triangular part: slvec0b = - B^T*xx */
2533: (*mat->B->ops->multtranspose)(mat->B,xx,mat->slvec0b);
2534:
2535: /* copy xx into slvec0a */
2536: VecGetArray(mat->slvec0,&ptr);
2537: VecGetArray(xx,&x);
2538: PetscMemcpy(ptr,x,bs*mbs*sizeof(MatScalar));
2539: VecRestoreArray(mat->slvec0,&ptr);
2541: VecScale(&mone,mat->slvec0);
2543: /* copy bb into slvec1a */
2544: VecGetArray(mat->slvec1,&ptr);
2545: VecGetArray(bb,&b);
2546: PetscMemcpy(ptr,b,bs*mbs*sizeof(MatScalar));
2547: VecRestoreArray(mat->slvec1,&ptr);
2549: /* set slvec1b = 0 */
2550: VecSet(&zero,mat->slvec1b);
2552: VecScatterBegin(mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD,mat->sMvctx);
2553: VecRestoreArray(xx,&x);
2554: VecRestoreArray(bb,&b);
2555: VecScatterEnd(mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD,mat->sMvctx);
2557: /* upper triangular part: bb1 = bb1 - B*x */
2558: (*mat->B->ops->multadd)(mat->B,mat->slvec1b,mat->slvec1a,bb1);
2559:
2560: /* local diagonal sweep */
2561: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
2562: }
2563: VecDestroy(bb1);
2564: } else {
2565: SETERRQ(PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format");
2566: }
2567: return(0);
2568: }
2572: PetscErrorCode MatRelax_MPISBAIJ_2comm(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
2573: {
2574: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
2576: PetscScalar mone=-1.0;
2577: Vec lvec1,bb1;
2578:
2580: if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
2581: if (matin->bs > 1)
2582: SETERRQ(PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented");
2584: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
2585: if ( flag & SOR_ZERO_INITIAL_GUESS ) {
2586: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
2587: its--;
2588: }
2590: VecDuplicate(mat->lvec,&lvec1);
2591: VecDuplicate(bb,&bb1);
2592: while (its--){
2593: VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
2594:
2595: /* lower diagonal part: bb1 = bb - B^T*xx */
2596: (*mat->B->ops->multtranspose)(mat->B,xx,lvec1);
2597: VecScale(&mone,lvec1);
2599: VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
2600: VecCopy(bb,bb1);
2601: VecScatterBegin(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);
2603: /* upper diagonal part: bb1 = bb1 - B*x */
2604: VecScale(&mone,mat->lvec);
2605: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb1,bb1);
2607: VecScatterEnd(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);
2608:
2609: /* diagonal sweep */
2610: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
2611: }
2612: VecDestroy(lvec1);
2613: VecDestroy(bb1);
2614: } else {
2615: SETERRQ(PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format");
2616: }
2617: return(0);
2618: }