Actual source code: mpisbaij.c
1: /*$Id: mpisbaij.c,v 1.61 2001/08/10 03:31:37 bsmith Exp $*/
3: #include src/mat/impls/baij/mpi/mpibaij.h
4: #include mpisbaij.h
5: #include src/mat/impls/sbaij/seq/sbaij.h
7: extern int MatSetUpMultiply_MPISBAIJ(Mat);
8: extern int MatSetUpMultiply_MPISBAIJ_2comm(Mat);
9: extern int DisAssemble_MPISBAIJ(Mat);
10: extern int MatIncreaseOverlap_MPISBAIJ(Mat,int,IS[],int);
11: extern int MatGetValues_SeqSBAIJ(Mat,int,const int[],int,const int[],PetscScalar []);
12: extern int MatGetValues_SeqBAIJ(Mat,int,const int[],int,const int[],PetscScalar []);
13: extern int MatSetValues_SeqSBAIJ(Mat,int,const int [],int,const int [],const PetscScalar [],InsertMode);
14: extern int MatSetValuesBlocked_SeqSBAIJ(Mat,int,const int[],int,const int[],const PetscScalar[],InsertMode);
15: extern int MatSetValuesBlocked_SeqBAIJ(Mat,int,const int[],int,const int[],const PetscScalar[],InsertMode);
16: extern int MatGetRow_SeqSBAIJ(Mat,int,int*,int**,PetscScalar**);
17: extern int MatRestoreRow_SeqSBAIJ(Mat,int,int*,int**,PetscScalar**);
18: extern int MatPrintHelp_SeqSBAIJ(Mat);
19: extern int MatZeroRows_SeqSBAIJ(Mat,IS,PetscScalar*);
20: extern int MatZeroRows_SeqBAIJ(Mat,IS,PetscScalar *);
21: extern int MatGetRowMax_MPISBAIJ(Mat,Vec);
22: extern int MatRelax_MPISBAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,int,int,Vec);
24: /* UGLY, ugly, ugly
25: When MatScalar == PetscScalar the function MatSetValuesBlocked_MPIBAIJ_MatScalar() does
26: not exist. Otherwise ..._MatScalar() takes matrix elements in single precision and
27: inserts them into the single precision data structure. The function MatSetValuesBlocked_MPIBAIJ()
28: converts the entries into single precision and then calls ..._MatScalar() to put them
29: into the single precision data structures.
30: */
31: #if defined(PETSC_USE_MAT_SINGLE)
32: extern int MatSetValuesBlocked_SeqSBAIJ_MatScalar(Mat,int,const int[],int,const int[],const MatScalar[],InsertMode);
33: extern int MatSetValues_MPISBAIJ_MatScalar(Mat,int,const int[],int,const int[],const MatScalar[],InsertMode);
34: extern int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat,int,const int[],int,const int[],const MatScalar[],InsertMode);
35: extern int MatSetValues_MPISBAIJ_HT_MatScalar(Mat,int,const int[],int,const int[],const MatScalar[],InsertMode);
36: extern int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat,int,const int[],int,const int[],const MatScalar[],InsertMode);
37: #else
38: #define MatSetValuesBlocked_SeqSBAIJ_MatScalar MatSetValuesBlocked_SeqSBAIJ
39: #define MatSetValues_MPISBAIJ_MatScalar MatSetValues_MPISBAIJ
40: #define MatSetValuesBlocked_MPISBAIJ_MatScalar MatSetValuesBlocked_MPISBAIJ
41: #define MatSetValues_MPISBAIJ_HT_MatScalar MatSetValues_MPISBAIJ_HT
42: #define MatSetValuesBlocked_MPISBAIJ_HT_MatScalar MatSetValuesBlocked_MPISBAIJ_HT
43: #endif
45: EXTERN_C_BEGIN
48: int MatStoreValues_MPISBAIJ(Mat mat)
49: {
50: Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data;
51: int ierr;
54: MatStoreValues(aij->A);
55: MatStoreValues(aij->B);
56: return(0);
57: }
58: EXTERN_C_END
60: EXTERN_C_BEGIN
63: int MatRetrieveValues_MPISBAIJ(Mat mat)
64: {
65: Mat_MPISBAIJ *aij = (Mat_MPISBAIJ *)mat->data;
66: int ierr;
69: MatRetrieveValues(aij->A);
70: MatRetrieveValues(aij->B);
71: return(0);
72: }
73: EXTERN_C_END
76: #define CHUNKSIZE 10
78: #define MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv) \
79: { \
80: \
81: brow = row/bs; \
82: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
83: rmax = aimax[brow]; nrow = ailen[brow]; \
84: bcol = col/bs; \
85: ridx = row % bs; cidx = col % bs; \
86: low = 0; high = nrow; \
87: while (high-low > 3) { \
88: t = (low+high)/2; \
89: if (rp[t] > bcol) high = t; \
90: else low = t; \
91: } \
92: for (_i=low; _i<high; _i++) { \
93: if (rp[_i] > bcol) break; \
94: if (rp[_i] == bcol) { \
95: bap = ap + bs2*_i + bs*cidx + ridx; \
96: if (addv == ADD_VALUES) *bap += value; \
97: else *bap = value; \
98: goto a_noinsert; \
99: } \
100: } \
101: if (a->nonew == 1) goto a_noinsert; \
102: else if (a->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) into matrix", row, col); \
103: if (nrow >= rmax) { \
104: /* there is no extra room in row, therefore enlarge */ \
105: int new_nz = ai[a->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
106: MatScalar *new_a; \
107: \
108: if (a->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) in the matrix", row, col); \
109: \
110: /* malloc new storage space */ \
111: len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(a->mbs+1)*sizeof(int); \
112: PetscMalloc(len,&new_a); \
113: new_j = (int*)(new_a + bs2*new_nz); \
114: new_i = new_j + new_nz; \
115: \
116: /* copy over old data into new slots */ \
117: for (ii=0; ii<brow+1; ii++) {new_i[ii] = ai[ii];} \
118: for (ii=brow+1; ii<a->mbs+1; ii++) {new_i[ii] = ai[ii]+CHUNKSIZE;} \
119: PetscMemcpy(new_j,aj,(ai[brow]+nrow)*sizeof(int)); \
120: len = (new_nz - CHUNKSIZE - ai[brow] - nrow); \
121: PetscMemcpy(new_j+ai[brow]+nrow+CHUNKSIZE,aj+ai[brow]+nrow,len*sizeof(int)); \
122: PetscMemcpy(new_a,aa,(ai[brow]+nrow)*bs2*sizeof(MatScalar)); \
123: PetscMemzero(new_a+bs2*(ai[brow]+nrow),bs2*CHUNKSIZE*sizeof(PetscScalar)); \
124: PetscMemcpy(new_a+bs2*(ai[brow]+nrow+CHUNKSIZE), \
125: aa+bs2*(ai[brow]+nrow),bs2*len*sizeof(MatScalar)); \
126: /* free up old matrix storage */ \
127: PetscFree(a->a); \
128: if (!a->singlemalloc) { \
129: PetscFree(a->i); \
130: PetscFree(a->j);\
131: } \
132: aa = a->a = new_a; ai = a->i = new_i; aj = a->j = new_j; \
133: a->singlemalloc = PETSC_TRUE; \
134: \
135: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
136: rmax = aimax[brow] = aimax[brow] + CHUNKSIZE; \
137: PetscLogObjectMemory(A,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \
138: a->maxnz += bs2*CHUNKSIZE; \
139: a->reallocs++; \
140: a->nz++; \
141: } \
142: N = nrow++ - 1; \
143: /* shift up all the later entries in this row */ \
144: for (ii=N; ii>=_i; ii--) { \
145: rp[ii+1] = rp[ii]; \
146: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
147: } \
148: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); } \
149: rp[_i] = bcol; \
150: ap[bs2*_i + bs*cidx + ridx] = value; \
151: a_noinsert:; \
152: ailen[brow] = nrow; \
153: }
154: #ifndef MatSetValues_SeqBAIJ_B_Private
155: #define MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv) \
156: { \
157: brow = row/bs; \
158: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
159: rmax = bimax[brow]; nrow = bilen[brow]; \
160: bcol = col/bs; \
161: ridx = row % bs; cidx = col % bs; \
162: low = 0; high = nrow; \
163: while (high-low > 3) { \
164: t = (low+high)/2; \
165: if (rp[t] > bcol) high = t; \
166: else low = t; \
167: } \
168: for (_i=low; _i<high; _i++) { \
169: if (rp[_i] > bcol) break; \
170: if (rp[_i] == bcol) { \
171: bap = ap + bs2*_i + bs*cidx + ridx; \
172: if (addv == ADD_VALUES) *bap += value; \
173: else *bap = value; \
174: goto b_noinsert; \
175: } \
176: } \
177: if (b->nonew == 1) goto b_noinsert; \
178: else if (b->nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) into matrix", row, col); \
179: if (nrow >= rmax) { \
180: /* there is no extra room in row, therefore enlarge */ \
181: int new_nz = bi[b->mbs] + CHUNKSIZE,len,*new_i,*new_j; \
182: MatScalar *new_a; \
183: \
184: if (b->nonew == -2) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%d, %d) in the matrix", row, col); \
185: \
186: /* malloc new storage space */ \
187: len = new_nz*(sizeof(int)+bs2*sizeof(MatScalar))+(b->mbs+1)*sizeof(int); \
188: PetscMalloc(len,&new_a); \
189: new_j = (int*)(new_a + bs2*new_nz); \
190: new_i = new_j + new_nz; \
191: \
192: /* copy over old data into new slots */ \
193: for (ii=0; ii<brow+1; ii++) {new_i[ii] = bi[ii];} \
194: for (ii=brow+1; ii<b->mbs+1; ii++) {new_i[ii] = bi[ii]+CHUNKSIZE;} \
195: PetscMemcpy(new_j,bj,(bi[brow]+nrow)*sizeof(int)); \
196: len = (new_nz - CHUNKSIZE - bi[brow] - nrow); \
197: PetscMemcpy(new_j+bi[brow]+nrow+CHUNKSIZE,bj+bi[brow]+nrow,len*sizeof(int)); \
198: PetscMemcpy(new_a,ba,(bi[brow]+nrow)*bs2*sizeof(MatScalar)); \
199: PetscMemzero(new_a+bs2*(bi[brow]+nrow),bs2*CHUNKSIZE*sizeof(MatScalar)); \
200: PetscMemcpy(new_a+bs2*(bi[brow]+nrow+CHUNKSIZE), \
201: ba+bs2*(bi[brow]+nrow),bs2*len*sizeof(MatScalar)); \
202: /* free up old matrix storage */ \
203: PetscFree(b->a); \
204: if (!b->singlemalloc) { \
205: PetscFree(b->i); \
206: PetscFree(b->j); \
207: } \
208: ba = b->a = new_a; bi = b->i = new_i; bj = b->j = new_j; \
209: b->singlemalloc = PETSC_TRUE; \
210: \
211: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
212: rmax = bimax[brow] = bimax[brow] + CHUNKSIZE; \
213: PetscLogObjectMemory(B,CHUNKSIZE*(sizeof(int) + bs2*sizeof(MatScalar))); \
214: b->maxnz += bs2*CHUNKSIZE; \
215: b->reallocs++; \
216: b->nz++; \
217: } \
218: N = nrow++ - 1; \
219: /* shift up all the later entries in this row */ \
220: for (ii=N; ii>=_i; ii--) { \
221: rp[ii+1] = rp[ii]; \
222: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
223: } \
224: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));} \
225: rp[_i] = bcol; \
226: ap[bs2*_i + bs*cidx + ridx] = value; \
227: b_noinsert:; \
228: bilen[brow] = nrow; \
229: }
230: #endif
232: #if defined(PETSC_USE_MAT_SINGLE)
235: int MatSetValues_MPISBAIJ(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
236: {
237: Mat_MPISBAIJ *b = (Mat_MPISBAIJ*)mat->data;
238: int ierr,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: int MatSetValuesBlocked_MPISBAIJ(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
259: {
260: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
261: int ierr,i,N = m*n*b->bs2;
262: MatScalar *vsingle;
265: if (N > b->setvalueslen) {
266: if (b->setvaluescopy) {PetscFree(b->setvaluescopy);}
267: PetscMalloc(N*sizeof(MatScalar),&b->setvaluescopy);
268: b->setvalueslen = N;
269: }
270: vsingle = b->setvaluescopy;
271: for (i=0; i<N; i++) {
272: vsingle[i] = v[i];
273: }
274: MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,m,im,n,in,vsingle,addv);
275: return(0);
276: }
280: int MatSetValues_MPISBAIJ_HT(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
281: {
282: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
283: int ierr,i,N = m*n;
284: MatScalar *vsingle;
287: SETERRQ(1,"Function not yet written for SBAIJ format");
288: /* return(0); */
289: }
293: int MatSetValuesBlocked_MPISBAIJ_HT(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
294: {
295: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)mat->data;
296: int ierr,i,N = m*n*b->bs2;
297: MatScalar *vsingle;
300: SETERRQ(1,"Function not yet written for SBAIJ format");
301: /* return(0); */
302: }
303: #endif
305: /* Only add/insert a(i,j) with i<=j (blocks).
306: Any a(i,j) with i>j input by user is ingored.
307: */
310: int MatSetValues_MPISBAIJ_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
311: {
312: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
313: MatScalar value;
314: PetscTruth roworiented = baij->roworiented;
315: int ierr,i,j,row,col;
316: int rstart_orig=baij->rstart_bs;
317: int rend_orig=baij->rend_bs,cstart_orig=baij->cstart_bs;
318: int cend_orig=baij->cend_bs,bs=baij->bs;
320: /* Some Variables required in the macro */
321: Mat A = baij->A;
322: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(A)->data;
323: int *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
324: MatScalar *aa=a->a;
326: Mat B = baij->B;
327: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data;
328: int *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
329: MatScalar *ba=b->a;
331: int *rp,ii,nrow,_i,rmax,N,brow,bcol;
332: int low,high,t,ridx,cidx,bs2=a->bs2;
333: MatScalar *ap,*bap;
335: /* for stash */
336: int n_loc, *in_loc=0;
337: MatScalar *v_loc=0;
341: if(!baij->donotstash){
342: PetscMalloc(n*sizeof(int),&in_loc);
343: PetscMalloc(n*sizeof(MatScalar),&v_loc);
344: }
346: for (i=0; i<m; i++) {
347: if (im[i] < 0) continue;
348: #if defined(PETSC_USE_BOPT_g)
349: if (im[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",im[i],mat->M-1);
350: #endif
351: if (im[i] >= rstart_orig && im[i] < rend_orig) { /* this processor entry */
352: row = im[i] - rstart_orig; /* local row index */
353: for (j=0; j<n; j++) {
354: if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */
355: if (in[j] >= cstart_orig && in[j] < cend_orig){ /* diag entry (A) */
356: col = in[j] - cstart_orig; /* local col index */
357: brow = row/bs; bcol = col/bs;
358: if (brow > bcol) continue; /* ignore lower triangular blocks of A */
359: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
360: MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv);
361: /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
362: } else if (in[j] < 0) continue;
363: #if defined(PETSC_USE_BOPT_g)
364: else if (in[j] >= mat->N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",in[j],mat->N-1);}
365: #endif
366: else { /* off-diag entry (B) */
367: if (mat->was_assembled) {
368: if (!baij->colmap) {
369: CreateColmap_MPIBAIJ_Private(mat);
370: }
371: #if defined (PETSC_USE_CTABLE)
372: PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
373: col = col - 1;
374: #else
375: col = baij->colmap[in[j]/bs] - 1;
376: #endif
377: if (col < 0 && !((Mat_SeqSBAIJ*)(baij->A->data))->nonew) {
378: DisAssemble_MPISBAIJ(mat);
379: col = in[j];
380: /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
381: B = baij->B;
382: b = (Mat_SeqBAIJ*)(B)->data;
383: bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
384: ba=b->a;
385: } else col += in[j]%bs;
386: } else col = in[j];
387: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
388: MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv);
389: /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
390: }
391: }
392: } else { /* off processor entry */
393: if (!baij->donotstash) {
394: n_loc = 0;
395: for (j=0; j<n; j++){
396: if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */
397: in_loc[n_loc] = in[j];
398: if (roworiented) {
399: v_loc[n_loc] = v[i*n+j];
400: } else {
401: v_loc[n_loc] = v[j*m+i];
402: }
403: n_loc++;
404: }
405: MatStashValuesRow_Private(&mat->stash,im[i],n_loc,in_loc,v_loc);
406: }
407: }
408: }
410: if(!baij->donotstash){
411: PetscFree(in_loc);
412: PetscFree(v_loc);
413: }
414: return(0);
415: }
419: int MatSetValuesBlocked_MPISBAIJ_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
420: {
421: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
422: const MatScalar *value;
423: MatScalar *barray=baij->barray;
424: PetscTruth roworiented = baij->roworiented;
425: int ierr,i,j,ii,jj,row,col,rstart=baij->rstart;
426: int rend=baij->rend,cstart=baij->cstart,stepval;
427: int cend=baij->cend,bs=baij->bs,bs2=baij->bs2;
430: if(!barray) {
431: PetscMalloc(bs2*sizeof(MatScalar),&barray);
432: baij->barray = barray;
433: }
435: if (roworiented) {
436: stepval = (n-1)*bs;
437: } else {
438: stepval = (m-1)*bs;
439: }
440: for (i=0; i<m; i++) {
441: if (im[i] < 0) continue;
442: #if defined(PETSC_USE_BOPT_g)
443: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %d max %d",im[i],baij->Mbs-1);
444: #endif
445: if (im[i] >= rstart && im[i] < rend) {
446: row = im[i] - rstart;
447: for (j=0; j<n; j++) {
448: /* If NumCol = 1 then a copy is not required */
449: if ((roworiented) && (n == 1)) {
450: barray = (MatScalar*) v + i*bs2;
451: } else if((!roworiented) && (m == 1)) {
452: barray = (MatScalar*) v + j*bs2;
453: } else { /* Here a copy is required */
454: if (roworiented) {
455: value = v + i*(stepval+bs)*bs + j*bs;
456: } else {
457: value = v + j*(stepval+bs)*bs + i*bs;
458: }
459: for (ii=0; ii<bs; ii++,value+=stepval) {
460: for (jj=0; jj<bs; jj++) {
461: *barray++ = *value++;
462: }
463: }
464: barray -=bs2;
465: }
466:
467: if (in[j] >= cstart && in[j] < cend){
468: col = in[j] - cstart;
469: MatSetValuesBlocked_SeqSBAIJ(baij->A,1,&row,1,&col,barray,addv);
470: }
471: else if (in[j] < 0) continue;
472: #if defined(PETSC_USE_BOPT_g)
473: else if (in[j] >= baij->Nbs) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %d max %d",in[j],baij->Nbs-1);}
474: #endif
475: else {
476: if (mat->was_assembled) {
477: if (!baij->colmap) {
478: CreateColmap_MPIBAIJ_Private(mat);
479: }
481: #if defined(PETSC_USE_BOPT_g)
482: #if defined (PETSC_USE_CTABLE)
483: { int data;
484: PetscTableFind(baij->colmap,in[j]+1,&data);
485: if ((data - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
486: }
487: #else
488: if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_ERR_PLIB,"Incorrect colmap");
489: #endif
490: #endif
491: #if defined (PETSC_USE_CTABLE)
492: PetscTableFind(baij->colmap,in[j]+1,&col);
493: col = (col - 1)/bs;
494: #else
495: col = (baij->colmap[in[j]] - 1)/bs;
496: #endif
497: if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
498: DisAssemble_MPISBAIJ(mat);
499: col = in[j];
500: }
501: }
502: else col = in[j];
503: MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
504: }
505: }
506: } else {
507: if (!baij->donotstash) {
508: if (roworiented) {
509: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
510: } else {
511: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
512: }
513: }
514: }
515: }
516: return(0);
517: }
519: #define HASH_KEY 0.6180339887
520: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(int)((size)*(tmp-(int)tmp)))
521: /* #define HASH(size,key) ((int)((size)*fmod(((key)*HASH_KEY),1))) */
522: /* #define HASH(size,key,tmp) ((int)((size)*fmod(((key)*HASH_KEY),1))) */
525: int MatSetValues_MPISBAIJ_HT_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
526: {
528: SETERRQ(1,"Function not yet written for SBAIJ format");
529: /* return(0); */
530: }
534: int MatSetValuesBlocked_MPISBAIJ_HT_MatScalar(Mat mat,int m,const int im[],int n,const int in[],const MatScalar v[],InsertMode addv)
535: {
537: SETERRQ(1,"Function not yet written for SBAIJ format");
538: /* return(0); */
539: }
543: int MatGetValues_MPISBAIJ(Mat mat,int m,const int idxm[],int n,const int idxn[],PetscScalar v[])
544: {
545: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
546: int bs=baij->bs,ierr,i,j,bsrstart = baij->rstart*bs,bsrend = baij->rend*bs;
547: int bscstart = baij->cstart*bs,bscend = baij->cend*bs,row,col,data;
550: for (i=0; i<m; i++) {
551: if (idxm[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %d",idxm[i]);
552: if (idxm[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",idxm[i],mat->M-1);
553: if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
554: row = idxm[i] - bsrstart;
555: for (j=0; j<n; j++) {
556: if (idxn[j] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column %d",idxn[j]);
557: if (idxn[j] >= mat->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",idxn[j],mat->N-1);
558: if (idxn[j] >= bscstart && idxn[j] < bscend){
559: col = idxn[j] - bscstart;
560: MatGetValues_SeqSBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
561: } else {
562: if (!baij->colmap) {
563: CreateColmap_MPIBAIJ_Private(mat);
564: }
565: #if defined (PETSC_USE_CTABLE)
566: PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
567: data --;
568: #else
569: data = baij->colmap[idxn[j]/bs]-1;
570: #endif
571: if((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
572: else {
573: col = data + idxn[j]%bs;
574: MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
575: }
576: }
577: }
578: } else {
579: SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
580: }
581: }
582: return(0);
583: }
587: int MatNorm_MPISBAIJ(Mat mat,NormType type,PetscReal *norm)
588: {
589: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
590: /* Mat_SeqSBAIJ *amat = (Mat_SeqSBAIJ*)baij->A->data; */
591: /* Mat_SeqBAIJ *bmat = (Mat_SeqBAIJ*)baij->B->data; */
592: int ierr;
593: PetscReal sum[2],*lnorm2;
596: if (baij->size == 1) {
597: MatNorm(baij->A,type,norm);
598: } else {
599: if (type == NORM_FROBENIUS) {
600: PetscMalloc(2*sizeof(PetscReal),&lnorm2);
601: MatNorm(baij->A,type,lnorm2);
602: *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2++; /* squar power of norm(A) */
603: MatNorm(baij->B,type,lnorm2);
604: *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2--; /* squar power of norm(B) */
605: /*
606: MPI_Comm_rank(mat->comm,&rank);
607: PetscSynchronizedPrintf(mat->comm,"[%d], lnorm2=%g, %g\n",rank,lnorm2[0],lnorm2[1]);
608: */
609: MPI_Allreduce(lnorm2,&sum,2,MPIU_REAL,MPI_SUM,mat->comm);
610: /*
611: PetscSynchronizedPrintf(mat->comm,"[%d], sum=%g, %g\n",rank,sum[0],sum[1]);
612: PetscSynchronizedFlush(mat->comm); */
613:
614: *norm = sqrt(sum[0] + 2*sum[1]);
615: PetscFree(lnorm2);
616: } else {
617: SETERRQ(PETSC_ERR_SUP,"No support for this norm yet");
618: }
619: }
620: return(0);
621: }
623: /*
624: Creates the hash table, and sets the table
625: This table is created only once.
626: If new entried need to be added to the matrix
627: then the hash table has to be destroyed and
628: recreated.
629: */
632: int MatCreateHashTable_MPISBAIJ_Private(Mat mat,PetscReal factor)
633: {
635: SETERRQ(1,"Function not yet written for SBAIJ format");
636: /* return(0); */
637: }
641: int MatAssemblyBegin_MPISBAIJ(Mat mat,MatAssemblyType mode)
642: {
643: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
644: int ierr,nstash,reallocs;
645: InsertMode addv;
648: if (baij->donotstash) {
649: return(0);
650: }
652: /* make sure all processors are either in INSERTMODE or ADDMODE */
653: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);
654: if (addv == (ADD_VALUES|INSERT_VALUES)) {
655: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
656: }
657: mat->insertmode = addv; /* in case this processor had no cache */
659: MatStashScatterBegin_Private(&mat->stash,baij->rowners_bs);
660: MatStashScatterBegin_Private(&mat->bstash,baij->rowners);
661: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
662: PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Stash has %d entries,uses %d mallocs.\n",nstash,reallocs);
663: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
664: PetscLogInfo(0,"MatAssemblyBegin_MPISBAIJ:Block-Stash has %d entries, uses %d mallocs.\n",nstash,reallocs);
665: return(0);
666: }
670: int MatAssemblyEnd_MPISBAIJ(Mat mat,MatAssemblyType mode)
671: {
672: Mat_MPISBAIJ *baij=(Mat_MPISBAIJ*)mat->data;
673: Mat_SeqSBAIJ *a=(Mat_SeqSBAIJ*)baij->A->data;
674: Mat_SeqBAIJ *b=(Mat_SeqBAIJ*)baij->B->data;
675: int i,j,rstart,ncols,n,ierr,flg,bs2=baij->bs2;
676: int *row,*col,other_disassembled;
677: PetscTruth r1,r2,r3;
678: MatScalar *val;
679: InsertMode addv = mat->insertmode;
683: if (!baij->donotstash) {
684: while (1) {
685: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
686: /*
687: PetscSynchronizedPrintf(mat->comm,"[%d]: in AssemblyEnd, stash, flg=%d\n",rank,flg);
688: PetscSynchronizedFlush(mat->comm);
689: */
690: if (!flg) break;
692: for (i=0; i<n;) {
693: /* Now identify the consecutive vals belonging to the same row */
694: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
695: if (j < n) ncols = j-i;
696: else ncols = n-i;
697: /* Now assemble all these values with a single function call */
698: MatSetValues_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i,addv);
699: i = j;
700: }
701: }
702: MatStashScatterEnd_Private(&mat->stash);
703: /* Now process the block-stash. Since the values are stashed column-oriented,
704: set the roworiented flag to column oriented, and after MatSetValues()
705: restore the original flags */
706: r1 = baij->roworiented;
707: r2 = a->roworiented;
708: r3 = b->roworiented;
709: baij->roworiented = PETSC_FALSE;
710: a->roworiented = PETSC_FALSE;
711: b->roworiented = PETSC_FALSE;
712: while (1) {
713: MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
714: if (!flg) break;
715:
716: for (i=0; i<n;) {
717: /* Now identify the consecutive vals belonging to the same row */
718: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
719: if (j < n) ncols = j-i;
720: else ncols = n-i;
721: MatSetValuesBlocked_MPISBAIJ_MatScalar(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
722: i = j;
723: }
724: }
725: MatStashScatterEnd_Private(&mat->bstash);
726: baij->roworiented = r1;
727: a->roworiented = r2;
728: b->roworiented = r3;
729: }
731: MatAssemblyBegin(baij->A,mode);
732: MatAssemblyEnd(baij->A,mode);
734: /* determine if any processor has disassembled, if so we must
735: also disassemble ourselfs, in order that we may reassemble. */
736: /*
737: if nonzero structure of submatrix B cannot change then we know that
738: no processor disassembled thus we can skip this stuff
739: */
740: if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
741: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);
742: if (mat->was_assembled && !other_disassembled) {
743: DisAssemble_MPISBAIJ(mat);
744: }
745: }
747: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
748: MatSetUpMultiply_MPISBAIJ(mat); /* setup Mvctx and sMvctx */
749: }
750: MatAssemblyBegin(baij->B,mode);
751: MatAssemblyEnd(baij->B,mode);
752:
753: #if defined(PETSC_USE_BOPT_g)
754: if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
755: PetscLogInfo(0,"MatAssemblyEnd_MPISBAIJ:Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);
756: baij->ht_total_ct = 0;
757: baij->ht_insert_ct = 0;
758: }
759: #endif
760: if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
761: MatCreateHashTable_MPISBAIJ_Private(mat,baij->ht_fact);
762: mat->ops->setvalues = MatSetValues_MPISBAIJ_HT;
763: mat->ops->setvaluesblocked = MatSetValuesBlocked_MPISBAIJ_HT;
764: }
766: if (baij->rowvalues) {
767: PetscFree(baij->rowvalues);
768: baij->rowvalues = 0;
769: }
771: return(0);
772: }
776: static int MatView_MPISBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
777: {
778: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
779: int ierr,bs = baij->bs,size = baij->size,rank = baij->rank;
780: PetscTruth isascii,isdraw;
781: PetscViewer sviewer;
782: PetscViewerFormat format;
785: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);
786: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
787: if (isascii) {
788: PetscViewerGetFormat(viewer,&format);
789: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
790: MatInfo info;
791: MPI_Comm_rank(mat->comm,&rank);
792: MatGetInfo(mat,MAT_LOCAL,&info);
793: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %d nz %d nz alloced %d bs %d mem %d\n",
794: rank,mat->m,(int)info.nz_used*bs,(int)info.nz_allocated*bs,
795: baij->bs,(int)info.memory);
796: MatGetInfo(baij->A,MAT_LOCAL,&info);
797: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);
798: MatGetInfo(baij->B,MAT_LOCAL,&info);
799: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %d \n",rank,(int)info.nz_used*bs);
800: PetscViewerFlush(viewer);
801: VecScatterView(baij->Mvctx,viewer);
802: return(0);
803: } else if (format == PETSC_VIEWER_ASCII_INFO) {
804: PetscViewerASCIIPrintf(viewer," block size is %d\n",bs);
805: return(0);
806: }
807: }
809: if (isdraw) {
810: PetscDraw draw;
811: PetscTruth isnull;
812: PetscViewerDrawGetDraw(viewer,0,&draw);
813: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
814: }
816: if (size == 1) {
817: PetscObjectSetName((PetscObject)baij->A,mat->name);
818: MatView(baij->A,viewer);
819: } else {
820: /* assemble the entire matrix onto first processor. */
821: Mat A;
822: Mat_SeqSBAIJ *Aloc;
823: Mat_SeqBAIJ *Bloc;
824: int M = mat->M,N = mat->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
825: MatScalar *a;
827: /* Should this be the same type as mat? */
828: if (!rank) {
829: MatCreate(mat->comm,M,N,M,N,&A);
830: } else {
831: MatCreate(mat->comm,0,0,M,N,&A);
832: }
833: MatSetType(A,MATMPISBAIJ);
834: MatMPISBAIJSetPreallocation(A,baij->bs,0,PETSC_NULL,0,PETSC_NULL);
835: PetscLogObjectParent(mat,A);
837: /* copy over the A part */
838: Aloc = (Mat_SeqSBAIJ*)baij->A->data;
839: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
840: PetscMalloc(bs*sizeof(int),&rvals);
842: for (i=0; i<mbs; i++) {
843: rvals[0] = bs*(baij->rstart + i);
844: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
845: for (j=ai[i]; j<ai[i+1]; j++) {
846: col = (baij->cstart+aj[j])*bs;
847: for (k=0; k<bs; k++) {
848: MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
849: col++; a += bs;
850: }
851: }
852: }
853: /* copy over the B part */
854: Bloc = (Mat_SeqBAIJ*)baij->B->data;
855: ai = Bloc->i; aj = Bloc->j; a = Bloc->a;
856: for (i=0; i<mbs; i++) {
857: rvals[0] = bs*(baij->rstart + i);
858: for (j=1; j<bs; j++) { rvals[j] = rvals[j-1] + 1; }
859: for (j=ai[i]; j<ai[i+1]; j++) {
860: col = baij->garray[aj[j]]*bs;
861: for (k=0; k<bs; k++) {
862: MatSetValues_MPISBAIJ_MatScalar(A,bs,rvals,1,&col,a,INSERT_VALUES);
863: col++; a += bs;
864: }
865: }
866: }
867: PetscFree(rvals);
868: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
869: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
870: /*
871: Everyone has to call to draw the matrix since the graphics waits are
872: synchronized across all processors that share the PetscDraw object
873: */
874: PetscViewerGetSingleton(viewer,&sviewer);
875: if (!rank) {
876: PetscObjectSetName((PetscObject)((Mat_MPISBAIJ*)(A->data))->A,mat->name);
877: MatView(((Mat_MPISBAIJ*)(A->data))->A,sviewer);
878: }
879: PetscViewerRestoreSingleton(viewer,&sviewer);
880: MatDestroy(A);
881: }
882: return(0);
883: }
887: int MatView_MPISBAIJ(Mat mat,PetscViewer viewer)
888: {
889: int ierr;
890: PetscTruth isascii,isdraw,issocket,isbinary;
893: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&isascii);
894: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
895: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
896: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
897: if (isascii || isdraw || issocket || isbinary) {
898: MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);
899: } else {
900: SETERRQ1(1,"Viewer type %s not supported by MPISBAIJ matrices",((PetscObject)viewer)->type_name);
901: }
902: return(0);
903: }
907: int MatDestroy_MPISBAIJ(Mat mat)
908: {
909: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
910: int ierr;
913: #if defined(PETSC_USE_LOG)
914: PetscLogObjectState((PetscObject)mat,"Rows=%d,Cols=%d",mat->M,mat->N);
915: #endif
916: MatStashDestroy_Private(&mat->stash);
917: MatStashDestroy_Private(&mat->bstash);
918: PetscFree(baij->rowners);
919: MatDestroy(baij->A);
920: MatDestroy(baij->B);
921: #if defined (PETSC_USE_CTABLE)
922: if (baij->colmap) {PetscTableDelete(baij->colmap);}
923: #else
924: if (baij->colmap) {PetscFree(baij->colmap);}
925: #endif
926: if (baij->garray) {PetscFree(baij->garray);}
927: if (baij->lvec) {VecDestroy(baij->lvec);}
928: if (baij->Mvctx) {VecScatterDestroy(baij->Mvctx);}
929: if (baij->slvec0) {
930: VecDestroy(baij->slvec0);
931: VecDestroy(baij->slvec0b);
932: }
933: if (baij->slvec1) {
934: VecDestroy(baij->slvec1);
935: VecDestroy(baij->slvec1a);
936: VecDestroy(baij->slvec1b);
937: }
938: if (baij->sMvctx) {VecScatterDestroy(baij->sMvctx);}
939: if (baij->rowvalues) {PetscFree(baij->rowvalues);}
940: if (baij->barray) {PetscFree(baij->barray);}
941: if (baij->hd) {PetscFree(baij->hd);}
942: #if defined(PETSC_USE_MAT_SINGLE)
943: if (baij->setvaluescopy) {PetscFree(baij->setvaluescopy);}
944: #endif
945: PetscFree(baij);
946: return(0);
947: }
951: int MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy)
952: {
953: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
954: int ierr,nt,mbs=a->mbs,bs=a->bs;
955: PetscScalar *x,*from,zero=0.0;
956:
958: VecGetLocalSize(xx,&nt);
959: if (nt != A->n) {
960: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
961: }
962: VecGetLocalSize(yy,&nt);
963: if (nt != A->m) {
964: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
965: }
967: /* diagonal part */
968: (*a->A->ops->mult)(a->A,xx,a->slvec1a);
969: VecSet(&zero,a->slvec1b);
971: /* subdiagonal part */
972: (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);
974: /* copy x into the vec slvec0 */
975: VecGetArray(a->slvec0,&from);
976: VecGetArray(xx,&x);
977: PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));
978: VecRestoreArray(a->slvec0,&from);
979:
980: VecScatterBegin(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
981: VecRestoreArray(xx,&x);
982: VecScatterEnd(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
983:
984: /* supperdiagonal part */
985: (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,yy);
986:
987: return(0);
988: }
992: int MatMult_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy)
993: {
994: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
995: int ierr,nt;
998: VecGetLocalSize(xx,&nt);
999: if (nt != A->n) {
1000: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1001: }
1002: VecGetLocalSize(yy,&nt);
1003: if (nt != A->m) {
1004: SETERRQ(PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1005: }
1007: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1008: /* do diagonal part */
1009: (*a->A->ops->mult)(a->A,xx,yy);
1010: /* do supperdiagonal part */
1011: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1012: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1013: /* do subdiagonal part */
1014: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1015: VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1016: VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1018: return(0);
1019: }
1023: int MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1024: {
1025: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1026: int ierr,mbs=a->mbs,bs=a->bs;
1027: PetscScalar *x,*from,zero=0.0;
1028:
1030: /*
1031: PetscSynchronizedPrintf(A->comm," MatMultAdd is called ...\n");
1032: PetscSynchronizedFlush(A->comm);
1033: */
1034: /* diagonal part */
1035: (*a->A->ops->multadd)(a->A,xx,yy,a->slvec1a);
1036: VecSet(&zero,a->slvec1b);
1038: /* subdiagonal part */
1039: (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);
1041: /* copy x into the vec slvec0 */
1042: VecGetArray(a->slvec0,&from);
1043: VecGetArray(xx,&x);
1044: PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));
1045: VecRestoreArray(a->slvec0,&from);
1046:
1047: VecScatterBegin(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
1048: VecRestoreArray(xx,&x);
1049: VecScatterEnd(a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD,a->sMvctx);
1050:
1051: /* supperdiagonal part */
1052: (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,zz);
1053:
1054: return(0);
1055: }
1059: int MatMultAdd_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy,Vec zz)
1060: {
1061: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1062: int ierr;
1065: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1066: /* do diagonal part */
1067: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1068: /* do supperdiagonal part */
1069: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
1070: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1072: /* do subdiagonal part */
1073: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1074: VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1075: VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
1077: return(0);
1078: }
1082: int MatMultTranspose_MPISBAIJ(Mat A,Vec xx,Vec yy)
1083: {
1087: MatMult(A,xx,yy);
1088: return(0);
1089: }
1093: int MatMultTransposeAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1094: {
1098: MatMultAdd(A,xx,yy,zz);
1099: return(0);
1100: }
1102: /*
1103: This only works correctly for square matrices where the subblock A->A is the
1104: diagonal block
1105: */
1108: int MatGetDiagonal_MPISBAIJ(Mat A,Vec v)
1109: {
1110: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1111: int ierr;
1114: /* if (a->M != a->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */
1115: MatGetDiagonal(a->A,v);
1116: return(0);
1117: }
1121: int MatScale_MPISBAIJ(const PetscScalar *aa,Mat A)
1122: {
1123: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1124: int ierr;
1127: MatScale(aa,a->A);
1128: MatScale(aa,a->B);
1129: return(0);
1130: }
1134: int MatGetRow_MPISBAIJ(Mat matin,int row,int *nz,int **idx,PetscScalar **v)
1135: {
1136: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
1137: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1138: int bs = mat->bs,bs2 = mat->bs2,i,ierr,*cworkA,*cworkB,**pcA,**pcB;
1139: int nztot,nzA,nzB,lrow,brstart = mat->rstart*bs,brend = mat->rend*bs;
1140: int *cmap,*idx_p,cstart = mat->cstart;
1143: if (mat->getrowactive == PETSC_TRUE) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1144: mat->getrowactive = PETSC_TRUE;
1146: if (!mat->rowvalues && (idx || v)) {
1147: /*
1148: allocate enough space to hold information from the longest row.
1149: */
1150: Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data;
1151: Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data;
1152: int max = 1,mbs = mat->mbs,tmp;
1153: for (i=0; i<mbs; i++) {
1154: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */
1155: if (max < tmp) { max = tmp; }
1156: }
1157: PetscMalloc(max*bs2*(sizeof(int)+sizeof(PetscScalar)),&mat->rowvalues);
1158: mat->rowindices = (int*)(mat->rowvalues + max*bs2);
1159: }
1160:
1161: if (row < brstart || row >= brend) SETERRQ(PETSC_ERR_SUP,"Only local rows")
1162: lrow = row - brstart; /* local row index */
1164: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1165: if (!v) {pvA = 0; pvB = 0;}
1166: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1167: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1168: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1169: nztot = nzA + nzB;
1171: cmap = mat->garray;
1172: if (v || idx) {
1173: if (nztot) {
1174: /* Sort by increasing column numbers, assuming A and B already sorted */
1175: int imark = -1;
1176: if (v) {
1177: *v = v_p = mat->rowvalues;
1178: for (i=0; i<nzB; i++) {
1179: if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1180: else break;
1181: }
1182: imark = i;
1183: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1184: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1185: }
1186: if (idx) {
1187: *idx = idx_p = mat->rowindices;
1188: if (imark > -1) {
1189: for (i=0; i<imark; i++) {
1190: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1191: }
1192: } else {
1193: for (i=0; i<nzB; i++) {
1194: if (cmap[cworkB[i]/bs] < cstart)
1195: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1196: else break;
1197: }
1198: imark = i;
1199: }
1200: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i];
1201: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1202: }
1203: } else {
1204: if (idx) *idx = 0;
1205: if (v) *v = 0;
1206: }
1207: }
1208: *nz = nztot;
1209: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1210: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1211: return(0);
1212: }
1216: int MatRestoreRow_MPISBAIJ(Mat mat,int row,int *nz,int **idx,PetscScalar **v)
1217: {
1218: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
1221: if (baij->getrowactive == PETSC_FALSE) {
1222: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1223: }
1224: baij->getrowactive = PETSC_FALSE;
1225: return(0);
1226: }
1230: int MatGetBlockSize_MPISBAIJ(Mat mat,int *bs)
1231: {
1232: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
1235: *bs = baij->bs;
1236: return(0);
1237: }
1241: int MatZeroEntries_MPISBAIJ(Mat A)
1242: {
1243: Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data;
1244: int ierr;
1247: MatZeroEntries(l->A);
1248: MatZeroEntries(l->B);
1249: return(0);
1250: }
1254: int MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1255: {
1256: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data;
1257: Mat A = a->A,B = a->B;
1258: int ierr;
1259: PetscReal isend[5],irecv[5];
1262: info->block_size = (PetscReal)a->bs;
1263: MatGetInfo(A,MAT_LOCAL,info);
1264: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1265: isend[3] = info->memory; isend[4] = info->mallocs;
1266: MatGetInfo(B,MAT_LOCAL,info);
1267: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1268: isend[3] += info->memory; isend[4] += info->mallocs;
1269: if (flag == MAT_LOCAL) {
1270: info->nz_used = isend[0];
1271: info->nz_allocated = isend[1];
1272: info->nz_unneeded = isend[2];
1273: info->memory = isend[3];
1274: info->mallocs = isend[4];
1275: } else if (flag == MAT_GLOBAL_MAX) {
1276: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);
1277: info->nz_used = irecv[0];
1278: info->nz_allocated = irecv[1];
1279: info->nz_unneeded = irecv[2];
1280: info->memory = irecv[3];
1281: info->mallocs = irecv[4];
1282: } else if (flag == MAT_GLOBAL_SUM) {
1283: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);
1284: info->nz_used = irecv[0];
1285: info->nz_allocated = irecv[1];
1286: info->nz_unneeded = irecv[2];
1287: info->memory = irecv[3];
1288: info->mallocs = irecv[4];
1289: } else {
1290: SETERRQ1(1,"Unknown MatInfoType argument %d",flag);
1291: }
1292: info->rows_global = (PetscReal)A->M;
1293: info->columns_global = (PetscReal)A->N;
1294: info->rows_local = (PetscReal)A->m;
1295: info->columns_local = (PetscReal)A->N;
1296: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1297: info->fill_ratio_needed = 0;
1298: info->factor_mallocs = 0;
1299: return(0);
1300: }
1304: int MatSetOption_MPISBAIJ(Mat A,MatOption op)
1305: {
1306: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1307: int ierr;
1310: switch (op) {
1311: case MAT_NO_NEW_NONZERO_LOCATIONS:
1312: case MAT_YES_NEW_NONZERO_LOCATIONS:
1313: case MAT_COLUMNS_UNSORTED:
1314: case MAT_COLUMNS_SORTED:
1315: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1316: case MAT_KEEP_ZEROED_ROWS:
1317: case MAT_NEW_NONZERO_LOCATION_ERR:
1318: MatSetOption(a->A,op);
1319: MatSetOption(a->B,op);
1320: break;
1321: case MAT_ROW_ORIENTED:
1322: a->roworiented = PETSC_TRUE;
1323: MatSetOption(a->A,op);
1324: MatSetOption(a->B,op);
1325: break;
1326: case MAT_ROWS_SORTED:
1327: case MAT_ROWS_UNSORTED:
1328: case MAT_YES_NEW_DIAGONALS:
1329: PetscLogInfo(A,"Info:MatSetOption_MPIBAIJ:Option ignored\n");
1330: break;
1331: case MAT_COLUMN_ORIENTED:
1332: a->roworiented = PETSC_FALSE;
1333: MatSetOption(a->A,op);
1334: MatSetOption(a->B,op);
1335: break;
1336: case MAT_IGNORE_OFF_PROC_ENTRIES:
1337: a->donotstash = PETSC_TRUE;
1338: break;
1339: case MAT_NO_NEW_DIAGONALS:
1340: SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1341: case MAT_USE_HASH_TABLE:
1342: a->ht_flag = PETSC_TRUE;
1343: break;
1344: case MAT_NOT_SYMMETRIC:
1345: case MAT_NOT_STRUCTURALLY_SYMMETRIC:
1346: case MAT_HERMITIAN:
1347: SETERRQ(PETSC_ERR_SUP,"Matrix must be symmetric");
1348: case MAT_SYMMETRIC:
1349: case MAT_STRUCTURALLY_SYMMETRIC:
1350: case MAT_NOT_HERMITIAN:
1351: case MAT_SYMMETRY_ETERNAL:
1352: case MAT_NOT_SYMMETRY_ETERNAL:
1353: break;
1354: default:
1355: SETERRQ(PETSC_ERR_SUP,"unknown option");
1356: }
1357: return(0);
1358: }
1362: int MatTranspose_MPISBAIJ(Mat A,Mat *B)
1363: {
1366: MatDuplicate(A,MAT_COPY_VALUES,B);
1367: return(0);
1368: }
1372: int MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr)
1373: {
1374: Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data;
1375: Mat a = baij->A,b = baij->B;
1376: int ierr,s1,s2,s3;
1379: if (ll != rr) {
1380: SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n");
1381: }
1382: MatGetLocalSize(mat,&s2,&s3);
1383: if (rr) {
1384: VecGetLocalSize(rr,&s1);
1385: if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1386: /* Overlap communication with computation. */
1387: VecScatterBegin(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1388: /*} if (ll) { */
1389: VecGetLocalSize(ll,&s1);
1390: if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1391: (*b->ops->diagonalscale)(b,ll,PETSC_NULL);
1392: /* } */
1393: /* scale the diagonal block */
1394: (*a->ops->diagonalscale)(a,ll,rr);
1396: /* if (rr) { */
1397: /* Do a scatter end and then right scale the off-diagonal block */
1398: VecScatterEnd(rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
1399: (*b->ops->diagonalscale)(b,PETSC_NULL,baij->lvec);
1400: }
1401:
1402: return(0);
1403: }
1407: int MatZeroRows_MPISBAIJ(Mat A,IS is,const PetscScalar *diag)
1408: {
1410: SETERRQ(PETSC_ERR_SUP,"No support for this function yet");
1411: }
1415: int MatPrintHelp_MPISBAIJ(Mat A)
1416: {
1417: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1418: MPI_Comm comm = A->comm;
1419: static int called = 0;
1420: int ierr;
1423: if (!a->rank) {
1424: MatPrintHelp_SeqSBAIJ(a->A);
1425: }
1426: if (called) {return(0);} else called = 1;
1427: (*PetscHelpPrintf)(comm," Options for MATMPISBAIJ matrix format (the defaults):\n");
1428: (*PetscHelpPrintf)(comm," -mat_use_hash_table <factor>: Use hashtable for efficient matrix assembly\n");
1429: return(0);
1430: }
1434: int MatSetUnfactored_MPISBAIJ(Mat A)
1435: {
1436: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
1437: int ierr;
1440: MatSetUnfactored(a->A);
1441: return(0);
1442: }
1444: static int MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat *);
1448: int MatEqual_MPISBAIJ(Mat A,Mat B,PetscTruth *flag)
1449: {
1450: Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data;
1451: Mat a,b,c,d;
1452: PetscTruth flg;
1453: int ierr;
1456: a = matA->A; b = matA->B;
1457: c = matB->A; d = matB->B;
1459: MatEqual(a,c,&flg);
1460: if (flg == PETSC_TRUE) {
1461: MatEqual(b,d,&flg);
1462: }
1463: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);
1464: return(0);
1465: }
1469: int MatSetUpPreallocation_MPISBAIJ(Mat A)
1470: {
1471: int ierr;
1474: MatMPISBAIJSetPreallocation(A,1,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1475: return(0);
1476: }
1480: int MatGetSubMatrices_MPISBAIJ(Mat A,int n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
1481: {
1482: int i,ierr;
1483: PetscTruth flg;
1485: for (i=0; i<n; i++) {
1486: ISEqual(irow[i],icol[i],&flg);
1487: if (!flg) {
1488: SETERRQ(1,"Can only get symmetric submatrix for MPISBAIJ matrices");
1489: }
1490: }
1491: MatGetSubMatrices_MPIBAIJ(A,n,irow,icol,scall,B);
1492: return(0);
1493: }
1494:
1496: /* -------------------------------------------------------------------*/
1497: static struct _MatOps MatOps_Values = {
1498: MatSetValues_MPISBAIJ,
1499: MatGetRow_MPISBAIJ,
1500: MatRestoreRow_MPISBAIJ,
1501: MatMult_MPISBAIJ,
1502: /* 4*/ MatMultAdd_MPISBAIJ,
1503: MatMultTranspose_MPISBAIJ,
1504: MatMultTransposeAdd_MPISBAIJ,
1505: 0,
1506: 0,
1507: 0,
1508: /*10*/ 0,
1509: 0,
1510: 0,
1511: MatRelax_MPISBAIJ,
1512: MatTranspose_MPISBAIJ,
1513: /*15*/ MatGetInfo_MPISBAIJ,
1514: MatEqual_MPISBAIJ,
1515: MatGetDiagonal_MPISBAIJ,
1516: MatDiagonalScale_MPISBAIJ,
1517: MatNorm_MPISBAIJ,
1518: /*20*/ MatAssemblyBegin_MPISBAIJ,
1519: MatAssemblyEnd_MPISBAIJ,
1520: 0,
1521: MatSetOption_MPISBAIJ,
1522: MatZeroEntries_MPISBAIJ,
1523: /*25*/ MatZeroRows_MPISBAIJ,
1524: 0,
1525: 0,
1526: 0,
1527: 0,
1528: /*30*/ MatSetUpPreallocation_MPISBAIJ,
1529: 0,
1530: 0,
1531: 0,
1532: 0,
1533: /*35*/ MatDuplicate_MPISBAIJ,
1534: 0,
1535: 0,
1536: 0,
1537: 0,
1538: /*40*/ 0,
1539: MatGetSubMatrices_MPISBAIJ,
1540: MatIncreaseOverlap_MPISBAIJ,
1541: MatGetValues_MPISBAIJ,
1542: 0,
1543: /*45*/ MatPrintHelp_MPISBAIJ,
1544: MatScale_MPISBAIJ,
1545: 0,
1546: 0,
1547: 0,
1548: /*50*/ MatGetBlockSize_MPISBAIJ,
1549: 0,
1550: 0,
1551: 0,
1552: 0,
1553: /*55*/ 0,
1554: 0,
1555: MatSetUnfactored_MPISBAIJ,
1556: 0,
1557: MatSetValuesBlocked_MPISBAIJ,
1558: /*60*/ 0,
1559: 0,
1560: 0,
1561: MatGetPetscMaps_Petsc,
1562: 0,
1563: /*65*/ 0,
1564: 0,
1565: 0,
1566: 0,
1567: 0,
1568: /*70*/ MatGetRowMax_MPISBAIJ,
1569: 0,
1570: 0,
1571: 0,
1572: 0,
1573: /*75*/ 0,
1574: 0,
1575: 0,
1576: 0,
1577: 0,
1578: /*80*/ 0,
1579: 0,
1580: 0,
1581: 0,
1582: /*85*/ MatLoad_MPISBAIJ
1583: };
1586: EXTERN_C_BEGIN
1589: int MatGetDiagonalBlock_MPISBAIJ(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *a)
1590: {
1592: *a = ((Mat_MPISBAIJ *)A->data)->A;
1593: *iscopy = PETSC_FALSE;
1594: return(0);
1595: }
1596: EXTERN_C_END
1598: EXTERN_C_BEGIN
1601: int MatMPISBAIJSetPreallocation_MPISBAIJ(Mat B,int bs,int d_nz,int *d_nnz,int o_nz,int *o_nnz)
1602: {
1603: Mat_MPISBAIJ *b;
1604: int ierr,i,mbs,Mbs;
1607: PetscOptionsGetInt(B->prefix,"-mat_block_size",&bs,PETSC_NULL);
1609: if (bs < 1) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive");
1610: if (d_nz == PETSC_DECIDE || d_nz == PETSC_DEFAULT) d_nz = 3;
1611: if (o_nz == PETSC_DECIDE || o_nz == PETSC_DEFAULT) o_nz = 1;
1612: if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %d",d_nz);
1613: if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %d",o_nz);
1614: if (d_nnz) {
1615: for (i=0; i<B->m/bs; i++) {
1616: 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]);
1617: }
1618: }
1619: if (o_nnz) {
1620: for (i=0; i<B->m/bs; i++) {
1621: 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]);
1622: }
1623: }
1624: B->preallocated = PETSC_TRUE;
1625: PetscSplitOwnershipBlock(B->comm,bs,&B->m,&B->M);
1626: PetscSplitOwnershipBlock(B->comm,bs,&B->n,&B->N);
1627: PetscMapCreateMPI(B->comm,B->m,B->M,&B->rmap);
1628: PetscMapCreateMPI(B->comm,B->m,B->M,&B->cmap);
1630: b = (Mat_MPISBAIJ*)B->data;
1631: mbs = B->m/bs;
1632: Mbs = B->M/bs;
1633: if (mbs*bs != B->m) {
1634: SETERRQ2(PETSC_ERR_ARG_SIZ,"No of local rows %d must be divisible by blocksize %d",B->m,bs);
1635: }
1637: b->bs = bs;
1638: b->bs2 = bs*bs;
1639: b->mbs = mbs;
1640: b->nbs = mbs;
1641: b->Mbs = Mbs;
1642: b->Nbs = Mbs;
1644: MPI_Allgather(&b->mbs,1,MPI_INT,b->rowners+1,1,MPI_INT,B->comm);
1645: b->rowners[0] = 0;
1646: for (i=2; i<=b->size; i++) {
1647: b->rowners[i] += b->rowners[i-1];
1648: }
1649: b->rstart = b->rowners[b->rank];
1650: b->rend = b->rowners[b->rank+1];
1651: b->cstart = b->rstart;
1652: b->cend = b->rend;
1653: for (i=0; i<=b->size; i++) {
1654: b->rowners_bs[i] = b->rowners[i]*bs;
1655: }
1656: b->rstart_bs = b-> rstart*bs;
1657: b->rend_bs = b->rend*bs;
1658:
1659: b->cstart_bs = b->cstart*bs;
1660: b->cend_bs = b->cend*bs;
1661:
1662: MatCreate(PETSC_COMM_SELF,B->m,B->m,B->m,B->m,&b->A);
1663: MatSetType(b->A,MATSEQSBAIJ);
1664: MatSeqSBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);
1665: PetscLogObjectParent(B,b->A);
1667: MatCreate(PETSC_COMM_SELF,B->m,B->M,B->m,B->M,&b->B);
1668: MatSetType(b->B,MATSEQBAIJ);
1669: MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);
1670: PetscLogObjectParent(B,b->B);
1672: /* build cache for off array entries formed */
1673: MatStashCreate_Private(B->comm,bs,&B->bstash);
1675: return(0);
1676: }
1677: EXTERN_C_END
1679: /*MC
1680: MATMPISBAIJ - MATMPISBAIJ = "mpisbaij" - A matrix type to be used for distributed symmetric sparse block matrices,
1681: based on block compressed sparse row format. Only the upper triangular portion of the matrix is stored.
1683: Options Database Keys:
1684: . -mat_type mpisbaij - sets the matrix type to "mpisbaij" during a call to MatSetFromOptions()
1686: Level: beginner
1688: .seealso: MatCreateMPISBAIJ
1689: M*/
1691: EXTERN_C_BEGIN
1694: int MatCreate_MPISBAIJ(Mat B)
1695: {
1696: Mat_MPISBAIJ *b;
1697: int ierr;
1698: PetscTruth flg;
1702: PetscNew(Mat_MPISBAIJ,&b);
1703: B->data = (void*)b;
1704: PetscMemzero(b,sizeof(Mat_MPISBAIJ));
1705: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
1707: B->ops->destroy = MatDestroy_MPISBAIJ;
1708: B->ops->view = MatView_MPISBAIJ;
1709: B->mapping = 0;
1710: B->factor = 0;
1711: B->assembled = PETSC_FALSE;
1713: B->insertmode = NOT_SET_VALUES;
1714: MPI_Comm_rank(B->comm,&b->rank);
1715: MPI_Comm_size(B->comm,&b->size);
1717: /* build local table of row and column ownerships */
1718: PetscMalloc(3*(b->size+2)*sizeof(int),&b->rowners);
1719: b->cowners = b->rowners + b->size + 2;
1720: b->rowners_bs = b->cowners + b->size + 2;
1721: PetscLogObjectMemory(B,3*(b->size+2)*sizeof(int)+sizeof(struct _p_Mat)+sizeof(Mat_MPISBAIJ));
1723: /* build cache for off array entries formed */
1724: MatStashCreate_Private(B->comm,1,&B->stash);
1725: b->donotstash = PETSC_FALSE;
1726: b->colmap = PETSC_NULL;
1727: b->garray = PETSC_NULL;
1728: b->roworiented = PETSC_TRUE;
1730: #if defined(PETSC_USE_MAT_SINGLE)
1731: /* stuff for MatSetValues_XXX in single precision */
1732: b->setvalueslen = 0;
1733: b->setvaluescopy = PETSC_NULL;
1734: #endif
1736: /* stuff used in block assembly */
1737: b->barray = 0;
1739: /* stuff used for matrix vector multiply */
1740: b->lvec = 0;
1741: b->Mvctx = 0;
1742: b->slvec0 = 0;
1743: b->slvec0b = 0;
1744: b->slvec1 = 0;
1745: b->slvec1a = 0;
1746: b->slvec1b = 0;
1747: b->sMvctx = 0;
1749: /* stuff for MatGetRow() */
1750: b->rowindices = 0;
1751: b->rowvalues = 0;
1752: b->getrowactive = PETSC_FALSE;
1754: /* hash table stuff */
1755: b->ht = 0;
1756: b->hd = 0;
1757: b->ht_size = 0;
1758: b->ht_flag = PETSC_FALSE;
1759: b->ht_fact = 0;
1760: b->ht_total_ct = 0;
1761: b->ht_insert_ct = 0;
1763: PetscOptionsHasName(B->prefix,"-mat_use_hash_table",&flg);
1764: if (flg) {
1765: PetscReal fact = 1.39;
1766: MatSetOption(B,MAT_USE_HASH_TABLE);
1767: PetscOptionsGetReal(B->prefix,"-mat_use_hash_table",&fact,PETSC_NULL);
1768: if (fact <= 1.0) fact = 1.39;
1769: MatMPIBAIJSetHashTableFactor(B,fact);
1770: PetscLogInfo(0,"MatCreateMPISBAIJ:Hash table Factor used %5.2f\n",fact);
1771: }
1772: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
1773: "MatStoreValues_MPISBAIJ",
1774: MatStoreValues_MPISBAIJ);
1775: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
1776: "MatRetrieveValues_MPISBAIJ",
1777: MatRetrieveValues_MPISBAIJ);
1778: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
1779: "MatGetDiagonalBlock_MPISBAIJ",
1780: MatGetDiagonalBlock_MPISBAIJ);
1781: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPISBAIJSetPreallocation_C",
1782: "MatMPISBAIJSetPreallocation_MPISBAIJ",
1783: MatMPISBAIJSetPreallocation_MPISBAIJ);
1784: return(0);
1785: }
1786: EXTERN_C_END
1788: /*MC
1789: MATSBAIJ - MATSBAIJ = "sbaij" - A matrix type to be used for symmetric block sparse matrices.
1791: This matrix type is identical to MATSEQSBAIJ when constructed with a single process communicator,
1792: and MATMPISBAIJ otherwise.
1794: Options Database Keys:
1795: . -mat_type sbaij - sets the matrix type to "sbaij" during a call to MatSetFromOptions()
1797: Level: beginner
1799: .seealso: MatCreateMPISBAIJ,MATSEQSBAIJ,MATMPISBAIJ
1800: M*/
1802: EXTERN_C_BEGIN
1805: int MatCreate_SBAIJ(Mat A) {
1806: int ierr,size;
1809: PetscObjectChangeTypeName((PetscObject)A,MATSBAIJ);
1810: MPI_Comm_size(A->comm,&size);
1811: if (size == 1) {
1812: MatSetType(A,MATSEQSBAIJ);
1813: } else {
1814: MatSetType(A,MATMPISBAIJ);
1815: }
1816: return(0);
1817: }
1818: EXTERN_C_END
1822: /*@C
1823: MatMPISBAIJSetPreallocation - For good matrix assembly performance
1824: the user should preallocate the matrix storage by setting the parameters
1825: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
1826: performance can be increased by more than a factor of 50.
1828: Collective on Mat
1830: Input Parameters:
1831: + A - the matrix
1832: . bs - size of blockk
1833: . d_nz - number of block nonzeros per block row in diagonal portion of local
1834: submatrix (same for all local rows)
1835: . d_nnz - array containing the number of block nonzeros in the various block rows
1836: in the upper triangular and diagonal part of the in diagonal portion of the local
1837: (possibly different for each block row) or PETSC_NULL. You must leave room
1838: for the diagonal entry even if it is zero.
1839: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
1840: submatrix (same for all local rows).
1841: - o_nnz - array containing the number of nonzeros in the various block rows of the
1842: off-diagonal portion of the local submatrix (possibly different for
1843: each block row) or PETSC_NULL.
1846: Options Database Keys:
1847: . -mat_no_unroll - uses code that does not unroll the loops in the
1848: block calculations (much slower)
1849: . -mat_block_size - size of the blocks to use
1851: Notes:
1853: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
1854: than it must be used on all processors that share the object for that argument.
1856: Storage Information:
1857: For a square global matrix we define each processor's diagonal portion
1858: to be its local rows and the corresponding columns (a square submatrix);
1859: each processor's off-diagonal portion encompasses the remainder of the
1860: local matrix (a rectangular submatrix).
1862: The user can specify preallocated storage for the diagonal part of
1863: the local submatrix with either d_nz or d_nnz (not both). Set
1864: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
1865: memory allocation. Likewise, specify preallocated storage for the
1866: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
1868: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
1869: the figure below we depict these three local rows and all columns (0-11).
1871: .vb
1872: 0 1 2 3 4 5 6 7 8 9 10 11
1873: -------------------
1874: row 3 | o o o d d d o o o o o o
1875: row 4 | o o o d d d o o o o o o
1876: row 5 | o o o d d d o o o o o o
1877: -------------------
1878: .ve
1879:
1880: Thus, any entries in the d locations are stored in the d (diagonal)
1881: submatrix, and any entries in the o locations are stored in the
1882: o (off-diagonal) submatrix. Note that the d matrix is stored in
1883: MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format.
1885: Now d_nz should indicate the number of block nonzeros per row in the upper triangular
1886: plus the diagonal part of the d matrix,
1887: and o_nz should indicate the number of block nonzeros per row in the o matrix.
1888: In general, for PDE problems in which most nonzeros are near the diagonal,
1889: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
1890: or you will get TERRIBLE performance; see the users' manual chapter on
1891: matrices.
1893: Level: intermediate
1895: .keywords: matrix, block, aij, compressed row, sparse, parallel
1897: .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
1898: @*/
1899: int MatMPISBAIJSetPreallocation(Mat B,int bs,int d_nz,const int d_nnz[],int o_nz,const int o_nnz[])
1900: {
1901: int ierr,(*f)(Mat,int,int,const int[],int,const int[]);
1904: PetscObjectQueryFunction((PetscObject)B,"MatMPISBAIJSetPreallocation_C",(void (**)(void))&f);
1905: if (f) {
1906: (*f)(B,bs,d_nz,d_nnz,o_nz,o_nnz);
1907: }
1908: return(0);
1909: }
1913: /*@C
1914: MatCreateMPISBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format
1915: (block compressed row). For good matrix assembly performance
1916: the user should preallocate the matrix storage by setting the parameters
1917: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
1918: performance can be increased by more than a factor of 50.
1920: Collective on MPI_Comm
1922: Input Parameters:
1923: + comm - MPI communicator
1924: . bs - size of blockk
1925: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
1926: This value should be the same as the local size used in creating the
1927: y vector for the matrix-vector product y = Ax.
1928: . n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
1929: This value should be the same as the local size used in creating the
1930: x vector for the matrix-vector product y = Ax.
1931: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
1932: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
1933: . d_nz - number of block nonzeros per block row in diagonal portion of local
1934: submatrix (same for all local rows)
1935: . d_nnz - array containing the number of block nonzeros in the various block rows
1936: in the upper triangular portion of the in diagonal portion of the local
1937: (possibly different for each block block row) or PETSC_NULL.
1938: You must leave room for the diagonal entry even if it is zero.
1939: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
1940: submatrix (same for all local rows).
1941: - o_nnz - array containing the number of nonzeros in the various block rows of the
1942: off-diagonal portion of the local submatrix (possibly different for
1943: each block row) or PETSC_NULL.
1945: Output Parameter:
1946: . A - the matrix
1948: Options Database Keys:
1949: . -mat_no_unroll - uses code that does not unroll the loops in the
1950: block calculations (much slower)
1951: . -mat_block_size - size of the blocks to use
1952: . -mat_mpi - use the parallel matrix data structures even on one processor
1953: (defaults to using SeqBAIJ format on one processor)
1955: Notes:
1956: The user MUST specify either the local or global matrix dimensions
1957: (possibly both).
1959: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
1960: than it must be used on all processors that share the object for that argument.
1962: Storage Information:
1963: For a square global matrix we define each processor's diagonal portion
1964: to be its local rows and the corresponding columns (a square submatrix);
1965: each processor's off-diagonal portion encompasses the remainder of the
1966: local matrix (a rectangular submatrix).
1968: The user can specify preallocated storage for the diagonal part of
1969: the local submatrix with either d_nz or d_nnz (not both). Set
1970: d_nz=PETSC_DEFAULT and d_nnz=PETSC_NULL for PETSc to control dynamic
1971: memory allocation. Likewise, specify preallocated storage for the
1972: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
1974: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
1975: the figure below we depict these three local rows and all columns (0-11).
1977: .vb
1978: 0 1 2 3 4 5 6 7 8 9 10 11
1979: -------------------
1980: row 3 | o o o d d d o o o o o o
1981: row 4 | o o o d d d o o o o o o
1982: row 5 | o o o d d d o o o o o o
1983: -------------------
1984: .ve
1985:
1986: Thus, any entries in the d locations are stored in the d (diagonal)
1987: submatrix, and any entries in the o locations are stored in the
1988: o (off-diagonal) submatrix. Note that the d matrix is stored in
1989: MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format.
1991: Now d_nz should indicate the number of block nonzeros per row in the upper triangular
1992: plus the diagonal part of the d matrix,
1993: and o_nz should indicate the number of block nonzeros per row in the o matrix.
1994: In general, for PDE problems in which most nonzeros are near the diagonal,
1995: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
1996: or you will get TERRIBLE performance; see the users' manual chapter on
1997: matrices.
1999: Level: intermediate
2001: .keywords: matrix, block, aij, compressed row, sparse, parallel
2003: .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateMPIBAIJ()
2004: @*/
2006: int MatCreateMPISBAIJ(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)
2007: {
2008: int ierr,size;
2011: MatCreate(comm,m,n,M,N,A);
2012: MPI_Comm_size(comm,&size);
2013: if (size > 1) {
2014: MatSetType(*A,MATMPISBAIJ);
2015: MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
2016: } else {
2017: MatSetType(*A,MATSEQSBAIJ);
2018: MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
2019: }
2020: return(0);
2021: }
2026: static int MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2027: {
2028: Mat mat;
2029: Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data;
2030: int ierr,len=0,nt,bs=oldmat->bs,mbs=oldmat->mbs;
2031: PetscScalar *array;
2034: *newmat = 0;
2035: MatCreate(matin->comm,matin->m,matin->n,matin->M,matin->N,&mat);
2036: MatSetType(mat,matin->type_name);
2038: /* PetscMemcpy(mat->ops,&MatOps_Values,sizeof(struct _MatOps)); */ /*-- cause error? */
2039: mat->factor = matin->factor;
2040: mat->preallocated = PETSC_TRUE;
2041: mat->assembled = PETSC_TRUE;
2042: mat->insertmode = NOT_SET_VALUES;
2044: a = (Mat_MPISBAIJ*)mat->data;
2045: a->bs = oldmat->bs;
2046: a->bs2 = oldmat->bs2;
2047: a->mbs = oldmat->mbs;
2048: a->nbs = oldmat->nbs;
2049: a->Mbs = oldmat->Mbs;
2050: a->Nbs = oldmat->Nbs;
2051:
2052: a->rstart = oldmat->rstart;
2053: a->rend = oldmat->rend;
2054: a->cstart = oldmat->cstart;
2055: a->cend = oldmat->cend;
2056: a->size = oldmat->size;
2057: a->rank = oldmat->rank;
2058: a->donotstash = oldmat->donotstash;
2059: a->roworiented = oldmat->roworiented;
2060: a->rowindices = 0;
2061: a->rowvalues = 0;
2062: a->getrowactive = PETSC_FALSE;
2063: a->barray = 0;
2064: a->rstart_bs = oldmat->rstart_bs;
2065: a->rend_bs = oldmat->rend_bs;
2066: a->cstart_bs = oldmat->cstart_bs;
2067: a->cend_bs = oldmat->cend_bs;
2069: /* hash table stuff */
2070: a->ht = 0;
2071: a->hd = 0;
2072: a->ht_size = 0;
2073: a->ht_flag = oldmat->ht_flag;
2074: a->ht_fact = oldmat->ht_fact;
2075: a->ht_total_ct = 0;
2076: a->ht_insert_ct = 0;
2077:
2078: PetscMemcpy(a->rowners,oldmat->rowners,3*(a->size+2)*sizeof(int));
2079: MatStashCreate_Private(matin->comm,1,&mat->stash);
2080: MatStashCreate_Private(matin->comm,oldmat->bs,&mat->bstash);
2081: if (oldmat->colmap) {
2082: #if defined (PETSC_USE_CTABLE)
2083: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2084: #else
2085: PetscMalloc((a->Nbs)*sizeof(int),&a->colmap);
2086: PetscLogObjectMemory(mat,(a->Nbs)*sizeof(int));
2087: PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(int));
2088: #endif
2089: } else a->colmap = 0;
2091: if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
2092: PetscMalloc(len*sizeof(int),&a->garray);
2093: PetscLogObjectMemory(mat,len*sizeof(int));
2094: PetscMemcpy(a->garray,oldmat->garray,len*sizeof(int));
2095: } else a->garray = 0;
2096:
2097: VecDuplicate(oldmat->lvec,&a->lvec);
2098: PetscLogObjectParent(mat,a->lvec);
2099: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2100: PetscLogObjectParent(mat,a->Mvctx);
2102: VecDuplicate(oldmat->slvec0,&a->slvec0);
2103: PetscLogObjectParent(mat,a->slvec0);
2104: VecDuplicate(oldmat->slvec1,&a->slvec1);
2105: PetscLogObjectParent(mat,a->slvec1);
2107: VecGetLocalSize(a->slvec1,&nt);
2108: VecGetArray(a->slvec1,&array);
2109: VecCreateSeqWithArray(PETSC_COMM_SELF,bs*mbs,array,&a->slvec1a);
2110: VecCreateSeqWithArray(PETSC_COMM_SELF,nt-bs*mbs,array+bs*mbs,&a->slvec1b);
2111: VecRestoreArray(a->slvec1,&array);
2112: VecGetArray(a->slvec0,&array);
2113: VecCreateSeqWithArray(PETSC_COMM_SELF,nt-bs*mbs,array+bs*mbs,&a->slvec0b);
2114: VecRestoreArray(a->slvec0,&array);
2115: PetscLogObjectParent(mat,a->slvec0);
2116: PetscLogObjectParent(mat,a->slvec1);
2117: PetscLogObjectParent(mat,a->slvec0b);
2118: PetscLogObjectParent(mat,a->slvec1a);
2119: PetscLogObjectParent(mat,a->slvec1b);
2121: /* VecScatterCopy(oldmat->sMvctx,&a->sMvctx); - not written yet, replaced by the lazy trick: */
2122: PetscObjectReference((PetscObject)oldmat->sMvctx);
2123: a->sMvctx = oldmat->sMvctx;
2124: PetscLogObjectParent(mat,a->sMvctx);
2126: MatDuplicate(oldmat->A,cpvalues,&a->A);
2127: PetscLogObjectParent(mat,a->A);
2128: MatDuplicate(oldmat->B,cpvalues,&a->B);
2129: PetscLogObjectParent(mat,a->B);
2130: PetscFListDuplicate(mat->qlist,&matin->qlist);
2131: *newmat = mat;
2132: return(0);
2133: }
2135: #include petscsys.h
2139: int MatLoad_MPISBAIJ(PetscViewer viewer,const MatType type,Mat *newmat)
2140: {
2141: Mat A;
2142: int i,nz,ierr,j,rstart,rend,fd;
2143: PetscScalar *vals,*buf;
2144: MPI_Comm comm = ((PetscObject)viewer)->comm;
2145: MPI_Status status;
2146: int header[4],rank,size,*rowlengths = 0,M,N,m,*rowners,*browners,maxnz,*cols;
2147: int *locrowlens,*sndcounts = 0,*procsnz = 0,jj,*mycols,*ibuf;
2148: int tag = ((PetscObject)viewer)->tag,bs=1,Mbs,mbs,extra_rows;
2149: int *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount;
2150: int dcount,kmax,k,nzcount,tmp;
2151:
2153: PetscOptionsGetInt(PETSC_NULL,"-matload_block_size",&bs,PETSC_NULL);
2155: MPI_Comm_size(comm,&size);
2156: MPI_Comm_rank(comm,&rank);
2157: if (!rank) {
2158: PetscViewerBinaryGetDescriptor(viewer,&fd);
2159: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
2160: if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2161: if (header[3] < 0) {
2162: SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ");
2163: }
2164: }
2166: MPI_Bcast(header+1,3,MPI_INT,0,comm);
2167: M = header[1]; N = header[2];
2169: if (M != N) SETERRQ(PETSC_ERR_SUP,"Can only do square matrices");
2171: /*
2172: This code adds extra rows to make sure the number of rows is
2173: divisible by the blocksize
2174: */
2175: Mbs = M/bs;
2176: extra_rows = bs - M + bs*(Mbs);
2177: if (extra_rows == bs) extra_rows = 0;
2178: else Mbs++;
2179: if (extra_rows &&!rank) {
2180: PetscLogInfo(0,"MatLoad_MPISBAIJ:Padding loaded matrix to match blocksize\n");
2181: }
2183: /* determine ownership of all rows */
2184: mbs = Mbs/size + ((Mbs % size) > rank);
2185: m = mbs*bs;
2186: PetscMalloc(2*(size+2)*sizeof(int),&rowners);
2187: browners = rowners + size + 1;
2188: MPI_Allgather(&mbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);
2189: rowners[0] = 0;
2190: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
2191: for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
2192: rstart = rowners[rank];
2193: rend = rowners[rank+1];
2194:
2195: /* distribute row lengths to all processors */
2196: PetscMalloc((rend-rstart)*bs*sizeof(int),&locrowlens);
2197: if (!rank) {
2198: PetscMalloc((M+extra_rows)*sizeof(int),&rowlengths);
2199: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
2200: for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1;
2201: PetscMalloc(size*sizeof(int),&sndcounts);
2202: for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i];
2203: MPI_Scatterv(rowlengths,sndcounts,browners,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);
2204: PetscFree(sndcounts);
2205: } else {
2206: MPI_Scatterv(0,0,0,MPI_INT,locrowlens,(rend-rstart)*bs,MPI_INT,0,comm);
2207: }
2208:
2209: if (!rank) { /* procs[0] */
2210: /* calculate the number of nonzeros on each processor */
2211: PetscMalloc(size*sizeof(int),&procsnz);
2212: PetscMemzero(procsnz,size*sizeof(int));
2213: for (i=0; i<size; i++) {
2214: for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) {
2215: procsnz[i] += rowlengths[j];
2216: }
2217: }
2218: PetscFree(rowlengths);
2219:
2220: /* determine max buffer needed and allocate it */
2221: maxnz = 0;
2222: for (i=0; i<size; i++) {
2223: maxnz = PetscMax(maxnz,procsnz[i]);
2224: }
2225: PetscMalloc(maxnz*sizeof(int),&cols);
2227: /* read in my part of the matrix column indices */
2228: nz = procsnz[0];
2229: PetscMalloc(nz*sizeof(int),&ibuf);
2230: mycols = ibuf;
2231: if (size == 1) nz -= extra_rows;
2232: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
2233: if (size == 1) for (i=0; i< extra_rows; i++) { mycols[nz+i] = M+i; }
2235: /* read in every ones (except the last) and ship off */
2236: for (i=1; i<size-1; i++) {
2237: nz = procsnz[i];
2238: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2239: MPI_Send(cols,nz,MPI_INT,i,tag,comm);
2240: }
2241: /* read in the stuff for the last proc */
2242: if (size != 1) {
2243: nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */
2244: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2245: for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
2246: MPI_Send(cols,nz+extra_rows,MPI_INT,size-1,tag,comm);
2247: }
2248: PetscFree(cols);
2249: } else { /* procs[i], i>0 */
2250: /* determine buffer space needed for message */
2251: nz = 0;
2252: for (i=0; i<m; i++) {
2253: nz += locrowlens[i];
2254: }
2255: PetscMalloc(nz*sizeof(int),&ibuf);
2256: mycols = ibuf;
2257: /* receive message of column indices*/
2258: MPI_Recv(mycols,nz,MPI_INT,0,tag,comm,&status);
2259: MPI_Get_count(&status,MPI_INT,&maxnz);
2260: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2261: }
2263: /* loop over local rows, determining number of off diagonal entries */
2264: PetscMalloc(2*(rend-rstart+1)*sizeof(int),&dlens);
2265: odlens = dlens + (rend-rstart);
2266: PetscMalloc(3*Mbs*sizeof(int),&mask);
2267: PetscMemzero(mask,3*Mbs*sizeof(int));
2268: masked1 = mask + Mbs;
2269: masked2 = masked1 + Mbs;
2270: rowcount = 0; nzcount = 0;
2271: for (i=0; i<mbs; i++) {
2272: dcount = 0;
2273: odcount = 0;
2274: for (j=0; j<bs; j++) {
2275: kmax = locrowlens[rowcount];
2276: for (k=0; k<kmax; k++) {
2277: tmp = mycols[nzcount++]/bs; /* block col. index */
2278: if (!mask[tmp]) {
2279: mask[tmp] = 1;
2280: if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */
2281: else masked1[dcount++] = tmp; /* entry in diag portion */
2282: }
2283: }
2284: rowcount++;
2285: }
2286:
2287: dlens[i] = dcount; /* d_nzz[i] */
2288: odlens[i] = odcount; /* o_nzz[i] */
2290: /* zero out the mask elements we set */
2291: for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
2292: for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
2293: }
2294:
2295: /* create our matrix */
2296: MatCreate(comm,m,m,PETSC_DETERMINE,PETSC_DETERMINE,&A);
2297: MatSetType(A,type);
2298: MatMPISBAIJSetPreallocation(A,bs,0,dlens,0,odlens);
2299: MatSetOption(A,MAT_COLUMNS_SORTED);
2300:
2301: if (!rank) {
2302: PetscMalloc(maxnz*sizeof(PetscScalar),&buf);
2303: /* read in my part of the matrix numerical values */
2304: nz = procsnz[0];
2305: vals = buf;
2306: mycols = ibuf;
2307: if (size == 1) nz -= extra_rows;
2308: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2309: if (size == 1) for (i=0; i< extra_rows; i++) { vals[nz+i] = 1.0; }
2311: /* insert into matrix */
2312: jj = rstart*bs;
2313: for (i=0; i<m; i++) {
2314: MatSetValues(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2315: mycols += locrowlens[i];
2316: vals += locrowlens[i];
2317: jj++;
2318: }
2320: /* read in other processors (except the last one) and ship out */
2321: for (i=1; i<size-1; i++) {
2322: nz = procsnz[i];
2323: vals = buf;
2324: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2325: MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);
2326: }
2327: /* the last proc */
2328: if (size != 1){
2329: nz = procsnz[i] - extra_rows;
2330: vals = buf;
2331: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2332: for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
2333: MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,A->tag,comm);
2334: }
2335: PetscFree(procsnz);
2337: } else {
2338: /* receive numeric values */
2339: PetscMalloc(nz*sizeof(PetscScalar),&buf);
2341: /* receive message of values*/
2342: vals = buf;
2343: mycols = ibuf;
2344: MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);
2345: MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2346: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2348: /* insert into matrix */
2349: jj = rstart*bs;
2350: for (i=0; i<m; i++) {
2351: MatSetValues_MPISBAIJ(A,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
2352: mycols += locrowlens[i];
2353: vals += locrowlens[i];
2354: jj++;
2355: }
2356: }
2358: PetscFree(locrowlens);
2359: PetscFree(buf);
2360: PetscFree(ibuf);
2361: PetscFree(rowners);
2362: PetscFree(dlens);
2363: PetscFree(mask);
2364: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2365: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2366: *newmat = A;
2367: return(0);
2368: }
2372: /*@
2373: MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.
2375: Input Parameters:
2376: . mat - the matrix
2377: . fact - factor
2379: Collective on Mat
2381: Level: advanced
2383: Notes:
2384: This can also be set by the command line option: -mat_use_hash_table fact
2386: .keywords: matrix, hashtable, factor, HT
2388: .seealso: MatSetOption()
2389: @*/
2390: int MatMPISBAIJSetHashTableFactor(Mat mat,PetscReal fact)
2391: {
2393: SETERRQ(1,"Function not yet written for SBAIJ format");
2394: /* return(0); */
2395: }
2399: int MatGetRowMax_MPISBAIJ(Mat A,Vec v)
2400: {
2401: Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data;
2402: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data;
2403: PetscReal atmp;
2404: PetscReal *work,*svalues,*rvalues;
2405: int ierr,i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol;
2406: int rank,size,*rowners_bs,dest,count,source;
2407: PetscScalar *va;
2408: MatScalar *ba;
2409: MPI_Status stat;
2412: MatGetRowMax(a->A,v);
2413: VecGetArray(v,&va);
2415: MPI_Comm_size(A->comm,&size);
2416: MPI_Comm_rank(A->comm,&rank);
2418: bs = a->bs;
2419: mbs = a->mbs;
2420: Mbs = a->Mbs;
2421: ba = b->a;
2422: bi = b->i;
2423: bj = b->j;
2424: /*
2425: PetscSynchronizedPrintf(A->comm,"[%d] M: %d, bs: %d, mbs: %d \n",rank,bs*Mbs,bs,mbs);
2426: PetscSynchronizedFlush(A->comm);
2427: */
2429: /* find ownerships */
2430: rowners_bs = a->rowners_bs;
2431: /*
2432: if (!rank){
2433: for (i=0; i<size+1; i++) PetscPrintf(PETSC_COMM_SELF," rowners_bs[%d]: %d\n",i,rowners_bs[i]);
2434: }
2435: */
2437: /* each proc creates an array to be distributed */
2438: PetscMalloc(bs*Mbs*sizeof(PetscReal),&work);
2439: PetscMemzero(work,bs*Mbs*sizeof(PetscReal));
2441: /* row_max for B */
2442: if (rank != size-1){
2443: for (i=0; i<mbs; i++) {
2444: ncols = bi[1] - bi[0]; bi++;
2445: brow = bs*i;
2446: for (j=0; j<ncols; j++){
2447: bcol = bs*(*bj);
2448: for (kcol=0; kcol<bs; kcol++){
2449: col = bcol + kcol; /* local col index */
2450: col += rowners_bs[rank+1]; /* global col index */
2451: /* PetscPrintf(PETSC_COMM_SELF,"[%d], col: %d\n",rank,col); */
2452: for (krow=0; krow<bs; krow++){
2453: atmp = PetscAbsScalar(*ba); ba++;
2454: row = brow + krow; /* local row index */
2455: /* printf("val[%d,%d]: %g\n",row,col,atmp); */
2456: if (PetscRealPart(va[row]) < atmp) va[row] = atmp;
2457: if (work[col] < atmp) work[col] = atmp;
2458: }
2459: }
2460: bj++;
2461: }
2462: }
2463: /*
2464: PetscPrintf(PETSC_COMM_SELF,"[%d], work: ",rank);
2465: for (i=0; i<bs*Mbs; i++) PetscPrintf(PETSC_COMM_SELF,"%g ",work[i]);
2466: PetscPrintf(PETSC_COMM_SELF,"[%d]: \n");
2467: */
2469: /* send values to its owners */
2470: for (dest=rank+1; dest<size; dest++){
2471: svalues = work + rowners_bs[dest];
2472: count = rowners_bs[dest+1]-rowners_bs[dest];
2473: MPI_Send(svalues,count,MPIU_REAL,dest,rank,A->comm);
2474: /*
2475: PetscSynchronizedPrintf(A->comm,"[%d] sends %d values to [%d]: %g, %g, %g, %g\n",rank,count,dest,svalues[0],svalues[1],svalues[2],svalues[3]);
2476: PetscSynchronizedFlush(A->comm);
2477: */
2478: }
2479: }
2480:
2481: /* receive values */
2482: if (rank){
2483: rvalues = work;
2484: count = rowners_bs[rank+1]-rowners_bs[rank];
2485: for (source=0; source<rank; source++){
2486: MPI_Recv(rvalues,count,MPIU_REAL,MPI_ANY_SOURCE,MPI_ANY_TAG,A->comm,&stat);
2487: /* process values */
2488: for (i=0; i<count; i++){
2489: if (PetscRealPart(va[i]) < rvalues[i]) va[i] = rvalues[i];
2490: }
2491: /*
2492: PetscSynchronizedPrintf(A->comm,"[%d] received %d values from [%d]: %g, %g, %g, %g \n",rank,count,stat.MPI_SOURCE,rvalues[0],rvalues[1],rvalues[2],rvalues[3]);
2493: PetscSynchronizedFlush(A->comm);
2494: */
2495: }
2496: }
2498: VecRestoreArray(v,&va);
2499: PetscFree(work);
2500: return(0);
2501: }
2505: int MatRelax_MPISBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,int its,int lits,Vec xx)
2506: {
2507: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
2508: int ierr,mbs=mat->mbs,bs=mat->bs;
2509: PetscScalar mone=-1.0,*x,*b,*ptr,zero=0.0;
2510: Vec bb1;
2511:
2513: if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %d and local its %d both positive",its,lits);
2514: if (bs > 1)
2515: SETERRQ(PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented");
2517: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
2518: if ( flag & SOR_ZERO_INITIAL_GUESS ) {
2519: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
2520: its--;
2521: }
2523: VecDuplicate(bb,&bb1);
2524: while (its--){
2525:
2526: /* lower triangular part: slvec0b = - B^T*xx */
2527: (*mat->B->ops->multtranspose)(mat->B,xx,mat->slvec0b);
2528:
2529: /* copy xx into slvec0a */
2530: VecGetArray(mat->slvec0,&ptr);
2531: VecGetArray(xx,&x);
2532: PetscMemcpy(ptr,x,bs*mbs*sizeof(MatScalar));
2533: VecRestoreArray(mat->slvec0,&ptr);
2535: VecScale(&mone,mat->slvec0);
2537: /* copy bb into slvec1a */
2538: VecGetArray(mat->slvec1,&ptr);
2539: VecGetArray(bb,&b);
2540: PetscMemcpy(ptr,b,bs*mbs*sizeof(MatScalar));
2541: VecRestoreArray(mat->slvec1,&ptr);
2543: /* set slvec1b = 0 */
2544: VecSet(&zero,mat->slvec1b);
2546: VecScatterBegin(mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD,mat->sMvctx);
2547: VecRestoreArray(xx,&x);
2548: VecRestoreArray(bb,&b);
2549: VecScatterEnd(mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD,mat->sMvctx);
2551: /* upper triangular part: bb1 = bb1 - B*x */
2552: (*mat->B->ops->multadd)(mat->B,mat->slvec1b,mat->slvec1a,bb1);
2553:
2554: /* local diagonal sweep */
2555: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
2556: }
2557: VecDestroy(bb1);
2558: } else {
2559: SETERRQ(PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format");
2560: }
2561: return(0);
2562: }
2566: int MatRelax_MPISBAIJ_2comm(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,int its,int lits,Vec xx)
2567: {
2568: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data;
2569: int ierr;
2570: PetscScalar mone=-1.0;
2571: Vec lvec1,bb1;
2572:
2574: if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %d and local its %d both positive",its,lits);
2575: if (mat->bs > 1)
2576: SETERRQ(PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented");
2578: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
2579: if ( flag & SOR_ZERO_INITIAL_GUESS ) {
2580: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
2581: its--;
2582: }
2584: VecDuplicate(mat->lvec,&lvec1);
2585: VecDuplicate(bb,&bb1);
2586: while (its--){
2587: VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
2588:
2589: /* lower diagonal part: bb1 = bb - B^T*xx */
2590: (*mat->B->ops->multtranspose)(mat->B,xx,lvec1);
2591: VecScale(&mone,lvec1);
2593: VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
2594: VecCopy(bb,bb1);
2595: VecScatterBegin(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);
2597: /* upper diagonal part: bb1 = bb1 - B*x */
2598: VecScale(&mone,mat->lvec);
2599: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb1,bb1);
2601: VecScatterEnd(lvec1,bb1,ADD_VALUES,SCATTER_REVERSE,mat->Mvctx);
2602:
2603: /* diagonal sweep */
2604: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
2605: }
2606: VecDestroy(lvec1);
2607: VecDestroy(bb1);
2608: } else {
2609: SETERRQ(PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format");
2610: }
2611: return(0);
2612: }