Actual source code: mpibaij.c
petsc-3.4.2 2013-07-02
2: #include <../src/mat/impls/baij/mpi/mpibaij.h> /*I "petscmat.h" I*/
3: #include <petscblaslapack.h>
5: extern PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat);
6: extern PetscErrorCode MatDisAssemble_MPIBAIJ(Mat);
7: extern PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],PetscScalar []);
8: extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],const PetscScalar [],InsertMode);
9: extern PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
10: extern PetscErrorCode MatGetRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
11: extern PetscErrorCode MatRestoreRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
12: extern PetscErrorCode MatZeroRows_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscScalar,Vec,Vec);
16: PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[])
17: {
18: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
20: PetscInt i,*idxb = 0;
21: PetscScalar *va,*vb;
22: Vec vtmp;
25: MatGetRowMaxAbs(a->A,v,idx);
26: VecGetArray(v,&va);
27: if (idx) {
28: for (i=0; i<A->rmap->n; i++) {
29: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
30: }
31: }
33: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
34: if (idx) {PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);}
35: MatGetRowMaxAbs(a->B,vtmp,idxb);
36: VecGetArray(vtmp,&vb);
38: for (i=0; i<A->rmap->n; i++) {
39: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
40: va[i] = vb[i];
41: if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs);
42: }
43: }
45: VecRestoreArray(v,&va);
46: VecRestoreArray(vtmp,&vb);
47: PetscFree(idxb);
48: VecDestroy(&vtmp);
49: return(0);
50: }
54: PetscErrorCode MatStoreValues_MPIBAIJ(Mat mat)
55: {
56: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data;
60: MatStoreValues(aij->A);
61: MatStoreValues(aij->B);
62: return(0);
63: }
67: PetscErrorCode MatRetrieveValues_MPIBAIJ(Mat mat)
68: {
69: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ*)mat->data;
73: MatRetrieveValues(aij->A);
74: MatRetrieveValues(aij->B);
75: return(0);
76: }
78: /*
79: Local utility routine that creates a mapping from the global column
80: number to the local number in the off-diagonal part of the local
81: storage of the matrix. This is done in a non scalable way since the
82: length of colmap equals the global matrix length.
83: */
86: PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat)
87: {
88: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
89: Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data;
91: PetscInt nbs = B->nbs,i,bs=mat->rmap->bs;
94: #if defined(PETSC_USE_CTABLE)
95: PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);
96: for (i=0; i<nbs; i++) {
97: PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);
98: }
99: #else
100: PetscMalloc((baij->Nbs+1)*sizeof(PetscInt),&baij->colmap);
101: PetscLogObjectMemory(mat,baij->Nbs*sizeof(PetscInt));
102: PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));
103: for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1;
104: #endif
105: return(0);
106: }
108: #define MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \
109: { \
110: \
111: brow = row/bs; \
112: rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
113: rmax = aimax[brow]; nrow = ailen[brow]; \
114: bcol = col/bs; \
115: ridx = row % bs; cidx = col % bs; \
116: low = 0; high = nrow; \
117: while (high-low > 3) { \
118: t = (low+high)/2; \
119: if (rp[t] > bcol) high = t; \
120: else low = t; \
121: } \
122: for (_i=low; _i<high; _i++) { \
123: if (rp[_i] > bcol) break; \
124: if (rp[_i] == bcol) { \
125: bap = ap + bs2*_i + bs*cidx + ridx; \
126: if (addv == ADD_VALUES) *bap += value; \
127: else *bap = value; \
128: goto a_noinsert; \
129: } \
130: } \
131: if (a->nonew == 1) goto a_noinsert; \
132: if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
133: MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \
134: N = nrow++ - 1; \
135: /* shift up all the later entries in this row */ \
136: for (ii=N; ii>=_i; ii--) { \
137: rp[ii+1] = rp[ii]; \
138: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
139: } \
140: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); } \
141: rp[_i] = bcol; \
142: ap[bs2*_i + bs*cidx + ridx] = value; \
143: a_noinsert:; \
144: ailen[brow] = nrow; \
145: }
147: #define MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \
148: { \
149: brow = row/bs; \
150: rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
151: rmax = bimax[brow]; nrow = bilen[brow]; \
152: bcol = col/bs; \
153: ridx = row % bs; cidx = col % bs; \
154: low = 0; high = nrow; \
155: while (high-low > 3) { \
156: t = (low+high)/2; \
157: if (rp[t] > bcol) high = t; \
158: else low = t; \
159: } \
160: for (_i=low; _i<high; _i++) { \
161: if (rp[_i] > bcol) break; \
162: if (rp[_i] == bcol) { \
163: bap = ap + bs2*_i + bs*cidx + ridx; \
164: if (addv == ADD_VALUES) *bap += value; \
165: else *bap = value; \
166: goto b_noinsert; \
167: } \
168: } \
169: if (b->nonew == 1) goto b_noinsert; \
170: if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
171: MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \
172: N = nrow++ - 1; \
173: /* shift up all the later entries in this row */ \
174: for (ii=N; ii>=_i; ii--) { \
175: rp[ii+1] = rp[ii]; \
176: PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
177: } \
178: if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));} \
179: rp[_i] = bcol; \
180: ap[bs2*_i + bs*cidx + ridx] = value; \
181: b_noinsert:; \
182: bilen[brow] = nrow; \
183: }
187: PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
188: {
189: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
190: MatScalar value;
191: PetscBool roworiented = baij->roworiented;
193: PetscInt i,j,row,col;
194: PetscInt rstart_orig=mat->rmap->rstart;
195: PetscInt rend_orig =mat->rmap->rend,cstart_orig=mat->cmap->rstart;
196: PetscInt cend_orig =mat->cmap->rend,bs=mat->rmap->bs;
198: /* Some Variables required in the macro */
199: Mat A = baij->A;
200: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)(A)->data;
201: PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
202: MatScalar *aa =a->a;
204: Mat B = baij->B;
205: Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data;
206: PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
207: MatScalar *ba =b->a;
209: PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol;
210: PetscInt low,high,t,ridx,cidx,bs2=a->bs2;
211: MatScalar *ap,*bap;
215: for (i=0; i<m; i++) {
216: if (im[i] < 0) continue;
217: #if defined(PETSC_USE_DEBUG)
218: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
219: #endif
220: if (im[i] >= rstart_orig && im[i] < rend_orig) {
221: row = im[i] - rstart_orig;
222: for (j=0; j<n; j++) {
223: if (in[j] >= cstart_orig && in[j] < cend_orig) {
224: col = in[j] - cstart_orig;
225: if (roworiented) value = v[i*n+j];
226: else value = v[i+j*m];
227: MatSetValues_SeqBAIJ_A_Private(row,col,value,addv);
228: /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
229: } else if (in[j] < 0) continue;
230: #if defined(PETSC_USE_DEBUG)
231: else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
232: #endif
233: else {
234: if (mat->was_assembled) {
235: if (!baij->colmap) {
236: MatCreateColmap_MPIBAIJ_Private(mat);
237: }
238: #if defined(PETSC_USE_CTABLE)
239: PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
240: col = col - 1;
241: #else
242: col = baij->colmap[in[j]/bs] - 1;
243: #endif
244: if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
245: MatDisAssemble_MPIBAIJ(mat);
246: col = in[j];
247: /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
248: B = baij->B;
249: b = (Mat_SeqBAIJ*)(B)->data;
250: bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
251: ba =b->a;
252: } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]);
253: else col += in[j]%bs;
254: } else col = in[j];
255: if (roworiented) value = v[i*n+j];
256: else value = v[i+j*m];
257: MatSetValues_SeqBAIJ_B_Private(row,col,value,addv);
258: /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
259: }
260: }
261: } else {
262: if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
263: if (!baij->donotstash) {
264: mat->assembled = PETSC_FALSE;
265: if (roworiented) {
266: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
267: } else {
268: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
269: }
270: }
271: }
272: }
273: return(0);
274: }
278: PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
279: {
280: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
281: const PetscScalar *value;
282: MatScalar *barray = baij->barray;
283: PetscBool roworiented = baij->roworiented;
284: PetscErrorCode ierr;
285: PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs;
286: PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval;
287: PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;
290: if (!barray) {
291: PetscMalloc(bs2*sizeof(MatScalar),&barray);
292: baij->barray = barray;
293: }
295: if (roworiented) stepval = (n-1)*bs;
296: else stepval = (m-1)*bs;
298: for (i=0; i<m; i++) {
299: if (im[i] < 0) continue;
300: #if defined(PETSC_USE_DEBUG)
301: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
302: #endif
303: if (im[i] >= rstart && im[i] < rend) {
304: row = im[i] - rstart;
305: for (j=0; j<n; j++) {
306: /* If NumCol = 1 then a copy is not required */
307: if ((roworiented) && (n == 1)) {
308: barray = (MatScalar*)v + i*bs2;
309: } else if ((!roworiented) && (m == 1)) {
310: barray = (MatScalar*)v + j*bs2;
311: } else { /* Here a copy is required */
312: if (roworiented) {
313: value = v + (i*(stepval+bs) + j)*bs;
314: } else {
315: value = v + (j*(stepval+bs) + i)*bs;
316: }
317: for (ii=0; ii<bs; ii++,value+=bs+stepval) {
318: for (jj=0; jj<bs; jj++) barray[jj] = value[jj];
319: barray += bs;
320: }
321: barray -= bs2;
322: }
324: if (in[j] >= cstart && in[j] < cend) {
325: col = in[j] - cstart;
326: MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
327: } else if (in[j] < 0) continue;
328: #if defined(PETSC_USE_DEBUG)
329: else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
330: #endif
331: else {
332: if (mat->was_assembled) {
333: if (!baij->colmap) {
334: MatCreateColmap_MPIBAIJ_Private(mat);
335: }
337: #if defined(PETSC_USE_DEBUG)
338: #if defined(PETSC_USE_CTABLE)
339: { PetscInt data;
340: PetscTableFind(baij->colmap,in[j]+1,&data);
341: if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
342: }
343: #else
344: if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
345: #endif
346: #endif
347: #if defined(PETSC_USE_CTABLE)
348: PetscTableFind(baij->colmap,in[j]+1,&col);
349: col = (col - 1)/bs;
350: #else
351: col = (baij->colmap[in[j]] - 1)/bs;
352: #endif
353: if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
354: MatDisAssemble_MPIBAIJ(mat);
355: col = in[j];
356: } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", bs*im[i], bs*in[j]);
357: } else col = in[j];
358: MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
359: }
360: }
361: } else {
362: if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
363: if (!baij->donotstash) {
364: if (roworiented) {
365: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
366: } else {
367: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
368: }
369: }
370: }
371: }
372: return(0);
373: }
375: #define HASH_KEY 0.6180339887
376: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp)))
377: /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
378: /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
381: PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
382: {
383: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
384: PetscBool roworiented = baij->roworiented;
386: PetscInt i,j,row,col;
387: PetscInt rstart_orig=mat->rmap->rstart;
388: PetscInt rend_orig =mat->rmap->rend,Nbs=baij->Nbs;
389: PetscInt h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx;
390: PetscReal tmp;
391: MatScalar **HD = baij->hd,value;
392: #if defined(PETSC_USE_DEBUG)
393: PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
394: #endif
398: for (i=0; i<m; i++) {
399: #if defined(PETSC_USE_DEBUG)
400: if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row");
401: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
402: #endif
403: row = im[i];
404: if (row >= rstart_orig && row < rend_orig) {
405: for (j=0; j<n; j++) {
406: col = in[j];
407: if (roworiented) value = v[i*n+j];
408: else value = v[i+j*m];
409: /* Look up PetscInto the Hash Table */
410: key = (row/bs)*Nbs+(col/bs)+1;
411: h1 = HASH(size,key,tmp);
414: idx = h1;
415: #if defined(PETSC_USE_DEBUG)
416: insert_ct++;
417: total_ct++;
418: if (HT[idx] != key) {
419: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
420: if (idx == size) {
421: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
422: if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
423: }
424: }
425: #else
426: if (HT[idx] != key) {
427: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
428: if (idx == size) {
429: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
430: if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
431: }
432: }
433: #endif
434: /* A HASH table entry is found, so insert the values at the correct address */
435: if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value;
436: else *(HD[idx]+ (col % bs)*bs + (row % bs)) = value;
437: }
438: } else if (!baij->donotstash) {
439: if (roworiented) {
440: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
441: } else {
442: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
443: }
444: }
445: }
446: #if defined(PETSC_USE_DEBUG)
447: baij->ht_total_ct = total_ct;
448: baij->ht_insert_ct = insert_ct;
449: #endif
450: return(0);
451: }
455: PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
456: {
457: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
458: PetscBool roworiented = baij->roworiented;
459: PetscErrorCode ierr;
460: PetscInt i,j,ii,jj,row,col;
461: PetscInt rstart=baij->rstartbs;
462: PetscInt rend =mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2;
463: PetscInt h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs;
464: PetscReal tmp;
465: MatScalar **HD = baij->hd,*baij_a;
466: const PetscScalar *v_t,*value;
467: #if defined(PETSC_USE_DEBUG)
468: PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
469: #endif
472: if (roworiented) stepval = (n-1)*bs;
473: else stepval = (m-1)*bs;
475: for (i=0; i<m; i++) {
476: #if defined(PETSC_USE_DEBUG)
477: if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]);
478: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1);
479: #endif
480: row = im[i];
481: v_t = v + i*nbs2;
482: if (row >= rstart && row < rend) {
483: for (j=0; j<n; j++) {
484: col = in[j];
486: /* Look up into the Hash Table */
487: key = row*Nbs+col+1;
488: h1 = HASH(size,key,tmp);
490: idx = h1;
491: #if defined(PETSC_USE_DEBUG)
492: total_ct++;
493: insert_ct++;
494: if (HT[idx] != key) {
495: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
496: if (idx == size) {
497: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
498: if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
499: }
500: }
501: #else
502: if (HT[idx] != key) {
503: for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
504: if (idx == size) {
505: for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
506: if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
507: }
508: }
509: #endif
510: baij_a = HD[idx];
511: if (roworiented) {
512: /*value = v + i*(stepval+bs)*bs + j*bs;*/
513: /* value = v + (i*(stepval+bs)+j)*bs; */
514: value = v_t;
515: v_t += bs;
516: if (addv == ADD_VALUES) {
517: for (ii=0; ii<bs; ii++,value+=stepval) {
518: for (jj=ii; jj<bs2; jj+=bs) {
519: baij_a[jj] += *value++;
520: }
521: }
522: } else {
523: for (ii=0; ii<bs; ii++,value+=stepval) {
524: for (jj=ii; jj<bs2; jj+=bs) {
525: baij_a[jj] = *value++;
526: }
527: }
528: }
529: } else {
530: value = v + j*(stepval+bs)*bs + i*bs;
531: if (addv == ADD_VALUES) {
532: for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
533: for (jj=0; jj<bs; jj++) {
534: baij_a[jj] += *value++;
535: }
536: }
537: } else {
538: for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
539: for (jj=0; jj<bs; jj++) {
540: baij_a[jj] = *value++;
541: }
542: }
543: }
544: }
545: }
546: } else {
547: if (!baij->donotstash) {
548: if (roworiented) {
549: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
550: } else {
551: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
552: }
553: }
554: }
555: }
556: #if defined(PETSC_USE_DEBUG)
557: baij->ht_total_ct = total_ct;
558: baij->ht_insert_ct = insert_ct;
559: #endif
560: return(0);
561: }
565: PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
566: {
567: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
569: PetscInt bs = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend;
570: PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data;
573: for (i=0; i<m; i++) {
574: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
575: if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
576: if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
577: row = idxm[i] - bsrstart;
578: for (j=0; j<n; j++) {
579: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
580: if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
581: if (idxn[j] >= bscstart && idxn[j] < bscend) {
582: col = idxn[j] - bscstart;
583: MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
584: } else {
585: if (!baij->colmap) {
586: MatCreateColmap_MPIBAIJ_Private(mat);
587: }
588: #if defined(PETSC_USE_CTABLE)
589: PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
590: data--;
591: #else
592: data = baij->colmap[idxn[j]/bs]-1;
593: #endif
594: if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
595: else {
596: col = data + idxn[j]%bs;
597: MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
598: }
599: }
600: }
601: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
602: }
603: return(0);
604: }
608: PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm)
609: {
610: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
611: Mat_SeqBAIJ *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data;
613: PetscInt i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col;
614: PetscReal sum = 0.0;
615: MatScalar *v;
618: if (baij->size == 1) {
619: MatNorm(baij->A,type,nrm);
620: } else {
621: if (type == NORM_FROBENIUS) {
622: v = amat->a;
623: nz = amat->nz*bs2;
624: for (i=0; i<nz; i++) {
625: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
626: }
627: v = bmat->a;
628: nz = bmat->nz*bs2;
629: for (i=0; i<nz; i++) {
630: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
631: }
632: MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
633: *nrm = PetscSqrtReal(*nrm);
634: } else if (type == NORM_1) { /* max column sum */
635: PetscReal *tmp,*tmp2;
636: PetscInt *jj,*garray=baij->garray,cstart=baij->rstartbs;
637: PetscMalloc2(mat->cmap->N,PetscReal,&tmp,mat->cmap->N,PetscReal,&tmp2);
638: PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));
639: v = amat->a; jj = amat->j;
640: for (i=0; i<amat->nz; i++) {
641: for (j=0; j<bs; j++) {
642: col = bs*(cstart + *jj) + j; /* column index */
643: for (row=0; row<bs; row++) {
644: tmp[col] += PetscAbsScalar(*v); v++;
645: }
646: }
647: jj++;
648: }
649: v = bmat->a; jj = bmat->j;
650: for (i=0; i<bmat->nz; i++) {
651: for (j=0; j<bs; j++) {
652: col = bs*garray[*jj] + j;
653: for (row=0; row<bs; row++) {
654: tmp[col] += PetscAbsScalar(*v); v++;
655: }
656: }
657: jj++;
658: }
659: MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
660: *nrm = 0.0;
661: for (j=0; j<mat->cmap->N; j++) {
662: if (tmp2[j] > *nrm) *nrm = tmp2[j];
663: }
664: PetscFree2(tmp,tmp2);
665: } else if (type == NORM_INFINITY) { /* max row sum */
666: PetscReal *sums;
667: PetscMalloc(bs*sizeof(PetscReal),&sums);
668: sum = 0.0;
669: for (j=0; j<amat->mbs; j++) {
670: for (row=0; row<bs; row++) sums[row] = 0.0;
671: v = amat->a + bs2*amat->i[j];
672: nz = amat->i[j+1]-amat->i[j];
673: for (i=0; i<nz; i++) {
674: for (col=0; col<bs; col++) {
675: for (row=0; row<bs; row++) {
676: sums[row] += PetscAbsScalar(*v); v++;
677: }
678: }
679: }
680: v = bmat->a + bs2*bmat->i[j];
681: nz = bmat->i[j+1]-bmat->i[j];
682: for (i=0; i<nz; i++) {
683: for (col=0; col<bs; col++) {
684: for (row=0; row<bs; row++) {
685: sums[row] += PetscAbsScalar(*v); v++;
686: }
687: }
688: }
689: for (row=0; row<bs; row++) {
690: if (sums[row] > sum) sum = sums[row];
691: }
692: }
693: MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
694: PetscFree(sums);
695: } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for this norm yet");
696: }
697: return(0);
698: }
700: /*
701: Creates the hash table, and sets the table
702: This table is created only once.
703: If new entried need to be added to the matrix
704: then the hash table has to be destroyed and
705: recreated.
706: */
709: PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor)
710: {
711: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
712: Mat A = baij->A,B=baij->B;
713: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ*)B->data;
714: PetscInt i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
716: PetscInt ht_size,bs2=baij->bs2,rstart=baij->rstartbs;
717: PetscInt cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs;
718: PetscInt *HT,key;
719: MatScalar **HD;
720: PetscReal tmp;
721: #if defined(PETSC_USE_INFO)
722: PetscInt ct=0,max=0;
723: #endif
726: if (baij->ht) return(0);
728: baij->ht_size = (PetscInt)(factor*nz);
729: ht_size = baij->ht_size;
731: /* Allocate Memory for Hash Table */
732: PetscMalloc2(ht_size,MatScalar*,&baij->hd,ht_size,PetscInt,&baij->ht);
733: PetscMemzero(baij->hd,ht_size*sizeof(MatScalar*));
734: PetscMemzero(baij->ht,ht_size*sizeof(PetscInt));
735: HD = baij->hd;
736: HT = baij->ht;
738: /* Loop Over A */
739: for (i=0; i<a->mbs; i++) {
740: for (j=ai[i]; j<ai[i+1]; j++) {
741: row = i+rstart;
742: col = aj[j]+cstart;
744: key = row*Nbs + col + 1;
745: h1 = HASH(ht_size,key,tmp);
746: for (k=0; k<ht_size; k++) {
747: if (!HT[(h1+k)%ht_size]) {
748: HT[(h1+k)%ht_size] = key;
749: HD[(h1+k)%ht_size] = a->a + j*bs2;
750: break;
751: #if defined(PETSC_USE_INFO)
752: } else {
753: ct++;
754: #endif
755: }
756: }
757: #if defined(PETSC_USE_INFO)
758: if (k> max) max = k;
759: #endif
760: }
761: }
762: /* Loop Over B */
763: for (i=0; i<b->mbs; i++) {
764: for (j=bi[i]; j<bi[i+1]; j++) {
765: row = i+rstart;
766: col = garray[bj[j]];
767: key = row*Nbs + col + 1;
768: h1 = HASH(ht_size,key,tmp);
769: for (k=0; k<ht_size; k++) {
770: if (!HT[(h1+k)%ht_size]) {
771: HT[(h1+k)%ht_size] = key;
772: HD[(h1+k)%ht_size] = b->a + j*bs2;
773: break;
774: #if defined(PETSC_USE_INFO)
775: } else {
776: ct++;
777: #endif
778: }
779: }
780: #if defined(PETSC_USE_INFO)
781: if (k> max) max = k;
782: #endif
783: }
784: }
786: /* Print Summary */
787: #if defined(PETSC_USE_INFO)
788: for (i=0,j=0; i<ht_size; i++) {
789: if (HT[i]) j++;
790: }
791: PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);
792: #endif
793: return(0);
794: }
798: PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode)
799: {
800: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
802: PetscInt nstash,reallocs;
803: InsertMode addv;
806: if (baij->donotstash || mat->nooffprocentries) return(0);
808: /* make sure all processors are either in INSERTMODE or ADDMODE */
809: MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));
810: if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
811: mat->insertmode = addv; /* in case this processor had no cache */
813: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
814: MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);
815: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
816: PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);
817: MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);
818: PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
819: return(0);
820: }
824: PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode)
825: {
826: Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data;
827: Mat_SeqBAIJ *a =(Mat_SeqBAIJ*)baij->A->data;
829: PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2;
830: PetscInt *row,*col;
831: PetscBool r1,r2,r3,other_disassembled;
832: MatScalar *val;
833: InsertMode addv = mat->insertmode;
834: PetscMPIInt n;
837: /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */
838: if (!baij->donotstash && !mat->nooffprocentries) {
839: while (1) {
840: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
841: if (!flg) break;
843: for (i=0; i<n;) {
844: /* Now identify the consecutive vals belonging to the same row */
845: for (j=i,rstart=row[j]; j<n; j++) {
846: if (row[j] != rstart) break;
847: }
848: if (j < n) ncols = j-i;
849: else ncols = n-i;
850: /* Now assemble all these values with a single function call */
851: MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
852: i = j;
853: }
854: }
855: MatStashScatterEnd_Private(&mat->stash);
856: /* Now process the block-stash. Since the values are stashed column-oriented,
857: set the roworiented flag to column oriented, and after MatSetValues()
858: restore the original flags */
859: r1 = baij->roworiented;
860: r2 = a->roworiented;
861: r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented;
863: baij->roworiented = PETSC_FALSE;
864: a->roworiented = PETSC_FALSE;
866: (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */
867: while (1) {
868: MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
869: if (!flg) break;
871: for (i=0; i<n;) {
872: /* Now identify the consecutive vals belonging to the same row */
873: for (j=i,rstart=row[j]; j<n; j++) {
874: if (row[j] != rstart) break;
875: }
876: if (j < n) ncols = j-i;
877: else ncols = n-i;
878: MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
879: i = j;
880: }
881: }
882: MatStashScatterEnd_Private(&mat->bstash);
884: baij->roworiented = r1;
885: a->roworiented = r2;
887: ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */
888: }
890: MatAssemblyBegin(baij->A,mode);
891: MatAssemblyEnd(baij->A,mode);
893: /* determine if any processor has disassembled, if so we must
894: also disassemble ourselfs, in order that we may reassemble. */
895: /*
896: if nonzero structure of submatrix B cannot change then we know that
897: no processor disassembled thus we can skip this stuff
898: */
899: if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
900: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
901: if (mat->was_assembled && !other_disassembled) {
902: MatDisAssemble_MPIBAIJ(mat);
903: }
904: }
906: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
907: MatSetUpMultiply_MPIBAIJ(mat);
908: }
909: MatSetOption(baij->B,MAT_CHECK_COMPRESSED_ROW,PETSC_FALSE);
910: MatAssemblyBegin(baij->B,mode);
911: MatAssemblyEnd(baij->B,mode);
913: #if defined(PETSC_USE_INFO)
914: if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
915: PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);
917: baij->ht_total_ct = 0;
918: baij->ht_insert_ct = 0;
919: }
920: #endif
921: if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
922: MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);
924: mat->ops->setvalues = MatSetValues_MPIBAIJ_HT;
925: mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT;
926: }
928: PetscFree2(baij->rowvalues,baij->rowindices);
930: baij->rowvalues = 0;
931: return(0);
932: }
934: #include <petscdraw.h>
937: static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
938: {
939: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
940: PetscErrorCode ierr;
941: PetscMPIInt size = baij->size,rank = baij->rank;
942: PetscInt bs = mat->rmap->bs;
943: PetscBool iascii,isdraw;
944: PetscViewer sviewer;
945: PetscViewerFormat format;
948: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
949: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
950: if (iascii) {
951: PetscViewerGetFormat(viewer,&format);
952: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
953: MatInfo info;
954: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
955: MatGetInfo(mat,MAT_LOCAL,&info);
956: PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
957: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n",
958: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);
959: MatGetInfo(baij->A,MAT_LOCAL,&info);
960: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
961: MatGetInfo(baij->B,MAT_LOCAL,&info);
962: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
963: PetscViewerFlush(viewer);
964: PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
965: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
966: VecScatterView(baij->Mvctx,viewer);
967: return(0);
968: } else if (format == PETSC_VIEWER_ASCII_INFO) {
969: PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);
970: return(0);
971: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
972: return(0);
973: }
974: }
976: if (isdraw) {
977: PetscDraw draw;
978: PetscBool isnull;
979: PetscViewerDrawGetDraw(viewer,0,&draw);
980: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
981: }
983: if (size == 1) {
984: PetscObjectSetName((PetscObject)baij->A,((PetscObject)mat)->name);
985: MatView(baij->A,viewer);
986: } else {
987: /* assemble the entire matrix onto first processor. */
988: Mat A;
989: Mat_SeqBAIJ *Aloc;
990: PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
991: MatScalar *a;
993: /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */
994: /* Perhaps this should be the type of mat? */
995: MatCreate(PetscObjectComm((PetscObject)mat),&A);
996: if (!rank) {
997: MatSetSizes(A,M,N,M,N);
998: } else {
999: MatSetSizes(A,0,0,M,N);
1000: }
1001: MatSetType(A,MATMPIBAIJ);
1002: MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);
1003: MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1004: PetscLogObjectParent(mat,A);
1006: /* copy over the A part */
1007: Aloc = (Mat_SeqBAIJ*)baij->A->data;
1008: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1009: PetscMalloc(bs*sizeof(PetscInt),&rvals);
1011: for (i=0; i<mbs; i++) {
1012: rvals[0] = bs*(baij->rstartbs + i);
1013: for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1014: for (j=ai[i]; j<ai[i+1]; j++) {
1015: col = (baij->cstartbs+aj[j])*bs;
1016: for (k=0; k<bs; k++) {
1017: MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1018: col++; a += bs;
1019: }
1020: }
1021: }
1022: /* copy over the B part */
1023: Aloc = (Mat_SeqBAIJ*)baij->B->data;
1024: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1025: for (i=0; i<mbs; i++) {
1026: rvals[0] = bs*(baij->rstartbs + i);
1027: for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1028: for (j=ai[i]; j<ai[i+1]; j++) {
1029: col = baij->garray[aj[j]]*bs;
1030: for (k=0; k<bs; k++) {
1031: MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1032: col++; a += bs;
1033: }
1034: }
1035: }
1036: PetscFree(rvals);
1037: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1038: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1039: /*
1040: Everyone has to call to draw the matrix since the graphics waits are
1041: synchronized across all processors that share the PetscDraw object
1042: */
1043: PetscViewerGetSingleton(viewer,&sviewer);
1044: if (!rank) {
1045: PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,((PetscObject)mat)->name);
1046: /* Set the type name to MATMPIBAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqBAIJ_ASCII()*/
1047: PetscStrcpy(((PetscObject)((Mat_MPIBAIJ*)(A->data))->A)->type_name,MATMPIBAIJ);
1048: MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);
1049: }
1050: PetscViewerRestoreSingleton(viewer,&sviewer);
1051: MatDestroy(&A);
1052: }
1053: return(0);
1054: }
1058: static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer)
1059: {
1060: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)mat->data;
1061: Mat_SeqBAIJ *A = (Mat_SeqBAIJ*)a->A->data;
1062: Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)a->B->data;
1064: PetscInt i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen;
1065: PetscInt *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll;
1066: int fd;
1067: PetscScalar *column_values;
1068: FILE *file;
1069: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
1070: PetscInt message_count,flowcontrolcount;
1073: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1074: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1075: nz = bs2*(A->nz + B->nz);
1076: rlen = mat->rmap->n;
1077: if (!rank) {
1078: header[0] = MAT_FILE_CLASSID;
1079: header[1] = mat->rmap->N;
1080: header[2] = mat->cmap->N;
1082: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1083: PetscViewerBinaryGetDescriptor(viewer,&fd);
1084: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1085: /* get largest number of rows any processor has */
1086: range = mat->rmap->range;
1087: for (i=1; i<size; i++) {
1088: rlen = PetscMax(rlen,range[i+1] - range[i]);
1089: }
1090: } else {
1091: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1092: }
1094: PetscMalloc((rlen/bs)*sizeof(PetscInt),&crow_lens);
1095: /* compute lengths of each row */
1096: for (i=0; i<a->mbs; i++) {
1097: crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1098: }
1099: /* store the row lengths to the file */
1100: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1101: if (!rank) {
1102: MPI_Status status;
1103: PetscMalloc(rlen*sizeof(PetscInt),&row_lens);
1104: rlen = (range[1] - range[0])/bs;
1105: for (i=0; i<rlen; i++) {
1106: for (j=0; j<bs; j++) {
1107: row_lens[i*bs+j] = bs*crow_lens[i];
1108: }
1109: }
1110: PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1111: for (i=1; i<size; i++) {
1112: rlen = (range[i+1] - range[i])/bs;
1113: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1114: MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1115: for (k=0; k<rlen; k++) {
1116: for (j=0; j<bs; j++) {
1117: row_lens[k*bs+j] = bs*crow_lens[k];
1118: }
1119: }
1120: PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1121: }
1122: PetscViewerFlowControlEndMaster(viewer,&message_count);
1123: PetscFree(row_lens);
1124: } else {
1125: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1126: MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1127: PetscViewerFlowControlEndWorker(viewer,&message_count);
1128: }
1129: PetscFree(crow_lens);
1131: /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the
1132: information needed to make it for each row from a block row. This does require more communication but still not more than
1133: the communication needed for the nonzero values */
1134: nzmax = nz; /* space a largest processor needs */
1135: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1136: PetscMalloc(nzmax*sizeof(PetscInt),&column_indices);
1137: cnt = 0;
1138: for (i=0; i<a->mbs; i++) {
1139: pcnt = cnt;
1140: for (j=B->i[i]; j<B->i[i+1]; j++) {
1141: if ((col = garray[B->j[j]]) > cstart) break;
1142: for (l=0; l<bs; l++) {
1143: column_indices[cnt++] = bs*col+l;
1144: }
1145: }
1146: for (k=A->i[i]; k<A->i[i+1]; k++) {
1147: for (l=0; l<bs; l++) {
1148: column_indices[cnt++] = bs*(A->j[k] + cstart)+l;
1149: }
1150: }
1151: for (; j<B->i[i+1]; j++) {
1152: for (l=0; l<bs; l++) {
1153: column_indices[cnt++] = bs*garray[B->j[j]]+l;
1154: }
1155: }
1156: len = cnt - pcnt;
1157: for (k=1; k<bs; k++) {
1158: PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));
1159: cnt += len;
1160: }
1161: }
1162: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1164: /* store the columns to the file */
1165: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1166: if (!rank) {
1167: MPI_Status status;
1168: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1169: for (i=1; i<size; i++) {
1170: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1171: MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1172: MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1173: PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);
1174: }
1175: PetscViewerFlowControlEndMaster(viewer,&message_count);
1176: } else {
1177: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1178: MPI_Send(&cnt,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1179: MPI_Send(column_indices,cnt,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1180: PetscViewerFlowControlEndWorker(viewer,&message_count);
1181: }
1182: PetscFree(column_indices);
1184: /* load up the numerical values */
1185: PetscMalloc(nzmax*sizeof(PetscScalar),&column_values);
1186: cnt = 0;
1187: for (i=0; i<a->mbs; i++) {
1188: rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]);
1189: for (j=B->i[i]; j<B->i[i+1]; j++) {
1190: if (garray[B->j[j]] > cstart) break;
1191: for (l=0; l<bs; l++) {
1192: for (ll=0; ll<bs; ll++) {
1193: column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1194: }
1195: }
1196: cnt += bs;
1197: }
1198: for (k=A->i[i]; k<A->i[i+1]; k++) {
1199: for (l=0; l<bs; l++) {
1200: for (ll=0; ll<bs; ll++) {
1201: column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll];
1202: }
1203: }
1204: cnt += bs;
1205: }
1206: for (; j<B->i[i+1]; j++) {
1207: for (l=0; l<bs; l++) {
1208: for (ll=0; ll<bs; ll++) {
1209: column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1210: }
1211: }
1212: cnt += bs;
1213: }
1214: cnt += (bs-1)*rlen;
1215: }
1216: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1218: /* store the column values to the file */
1219: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1220: if (!rank) {
1221: MPI_Status status;
1222: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1223: for (i=1; i<size; i++) {
1224: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1225: MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1226: MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat),&status);
1227: PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);
1228: }
1229: PetscViewerFlowControlEndMaster(viewer,&message_count);
1230: } else {
1231: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1232: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1233: MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1234: PetscViewerFlowControlEndWorker(viewer,&message_count);
1235: }
1236: PetscFree(column_values);
1238: PetscViewerBinaryGetInfoPointer(viewer,&file);
1239: if (file) {
1240: fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs);
1241: }
1242: return(0);
1243: }
1247: PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer)
1248: {
1250: PetscBool iascii,isdraw,issocket,isbinary;
1253: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1254: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1255: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1256: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1257: if (iascii || isdraw || issocket) {
1258: MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);
1259: } else if (isbinary) {
1260: MatView_MPIBAIJ_Binary(mat,viewer);
1261: }
1262: return(0);
1263: }
1267: PetscErrorCode MatDestroy_MPIBAIJ(Mat mat)
1268: {
1269: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1273: #if defined(PETSC_USE_LOG)
1274: PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N);
1275: #endif
1276: MatStashDestroy_Private(&mat->stash);
1277: MatStashDestroy_Private(&mat->bstash);
1278: MatDestroy(&baij->A);
1279: MatDestroy(&baij->B);
1280: #if defined(PETSC_USE_CTABLE)
1281: PetscTableDestroy(&baij->colmap);
1282: #else
1283: PetscFree(baij->colmap);
1284: #endif
1285: PetscFree(baij->garray);
1286: VecDestroy(&baij->lvec);
1287: VecScatterDestroy(&baij->Mvctx);
1288: PetscFree2(baij->rowvalues,baij->rowindices);
1289: PetscFree(baij->barray);
1290: PetscFree2(baij->hd,baij->ht);
1291: PetscFree(baij->rangebs);
1292: PetscFree(mat->data);
1294: PetscObjectChangeTypeName((PetscObject)mat,0);
1295: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1296: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1297: PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);
1298: PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C",NULL);
1299: PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C",NULL);
1300: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1301: PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C",NULL);
1302: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C",NULL);
1303: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C",NULL);
1304: return(0);
1305: }
1309: PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy)
1310: {
1311: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1313: PetscInt nt;
1316: VecGetLocalSize(xx,&nt);
1317: if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1318: VecGetLocalSize(yy,&nt);
1319: if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1320: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1321: (*a->A->ops->mult)(a->A,xx,yy);
1322: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1323: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1324: return(0);
1325: }
1329: PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1330: {
1331: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1335: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1336: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1337: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1338: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1339: return(0);
1340: }
1344: PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy)
1345: {
1346: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1348: PetscBool merged;
1351: VecScatterGetMerged(a->Mvctx,&merged);
1352: /* do nondiagonal part */
1353: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1354: if (!merged) {
1355: /* send it on its way */
1356: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1357: /* do local part */
1358: (*a->A->ops->multtranspose)(a->A,xx,yy);
1359: /* receive remote parts: note this assumes the values are not actually */
1360: /* inserted in yy until the next line */
1361: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1362: } else {
1363: /* do local part */
1364: (*a->A->ops->multtranspose)(a->A,xx,yy);
1365: /* send it on its way */
1366: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1367: /* values actually were received in the Begin() but we need to call this nop */
1368: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1369: }
1370: return(0);
1371: }
1375: PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1376: {
1377: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1381: /* do nondiagonal part */
1382: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1383: /* send it on its way */
1384: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1385: /* do local part */
1386: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1387: /* receive remote parts: note this assumes the values are not actually */
1388: /* inserted in yy until the next line, which is true for my implementation*/
1389: /* but is not perhaps always true. */
1390: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1391: return(0);
1392: }
1394: /*
1395: This only works correctly for square matrices where the subblock A->A is the
1396: diagonal block
1397: */
1400: PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v)
1401: {
1402: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1406: if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1407: MatGetDiagonal(a->A,v);
1408: return(0);
1409: }
1413: PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa)
1414: {
1415: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1419: MatScale(a->A,aa);
1420: MatScale(a->B,aa);
1421: return(0);
1422: }
1426: PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1427: {
1428: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data;
1429: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1431: PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB;
1432: PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend;
1433: PetscInt *cmap,*idx_p,cstart = mat->cstartbs;
1436: if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows");
1437: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1438: mat->getrowactive = PETSC_TRUE;
1440: if (!mat->rowvalues && (idx || v)) {
1441: /*
1442: allocate enough space to hold information from the longest row.
1443: */
1444: Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data;
1445: PetscInt max = 1,mbs = mat->mbs,tmp;
1446: for (i=0; i<mbs; i++) {
1447: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1448: if (max < tmp) max = tmp;
1449: }
1450: PetscMalloc2(max*bs2,PetscScalar,&mat->rowvalues,max*bs2,PetscInt,&mat->rowindices);
1451: }
1452: lrow = row - brstart;
1454: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1455: if (!v) {pvA = 0; pvB = 0;}
1456: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1457: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1458: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1459: nztot = nzA + nzB;
1461: cmap = mat->garray;
1462: if (v || idx) {
1463: if (nztot) {
1464: /* Sort by increasing column numbers, assuming A and B already sorted */
1465: PetscInt imark = -1;
1466: if (v) {
1467: *v = v_p = mat->rowvalues;
1468: for (i=0; i<nzB; i++) {
1469: if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1470: else break;
1471: }
1472: imark = i;
1473: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1474: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1475: }
1476: if (idx) {
1477: *idx = idx_p = mat->rowindices;
1478: if (imark > -1) {
1479: for (i=0; i<imark; i++) {
1480: idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1481: }
1482: } else {
1483: for (i=0; i<nzB; i++) {
1484: if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1485: else break;
1486: }
1487: imark = i;
1488: }
1489: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i];
1490: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1491: }
1492: } else {
1493: if (idx) *idx = 0;
1494: if (v) *v = 0;
1495: }
1496: }
1497: *nz = nztot;
1498: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1499: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1500: return(0);
1501: }
1505: PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1506: {
1507: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1510: if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1511: baij->getrowactive = PETSC_FALSE;
1512: return(0);
1513: }
1517: PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A)
1518: {
1519: Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data;
1523: MatZeroEntries(l->A);
1524: MatZeroEntries(l->B);
1525: return(0);
1526: }
1530: PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1531: {
1532: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)matin->data;
1533: Mat A = a->A,B = a->B;
1535: PetscReal isend[5],irecv[5];
1538: info->block_size = (PetscReal)matin->rmap->bs;
1540: MatGetInfo(A,MAT_LOCAL,info);
1542: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1543: isend[3] = info->memory; isend[4] = info->mallocs;
1545: MatGetInfo(B,MAT_LOCAL,info);
1547: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1548: isend[3] += info->memory; isend[4] += info->mallocs;
1550: if (flag == MAT_LOCAL) {
1551: info->nz_used = isend[0];
1552: info->nz_allocated = isend[1];
1553: info->nz_unneeded = isend[2];
1554: info->memory = isend[3];
1555: info->mallocs = isend[4];
1556: } else if (flag == MAT_GLOBAL_MAX) {
1557: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));
1559: info->nz_used = irecv[0];
1560: info->nz_allocated = irecv[1];
1561: info->nz_unneeded = irecv[2];
1562: info->memory = irecv[3];
1563: info->mallocs = irecv[4];
1564: } else if (flag == MAT_GLOBAL_SUM) {
1565: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));
1567: info->nz_used = irecv[0];
1568: info->nz_allocated = irecv[1];
1569: info->nz_unneeded = irecv[2];
1570: info->memory = irecv[3];
1571: info->mallocs = irecv[4];
1572: } else SETERRQ1(PetscObjectComm((PetscObject)matin),PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag);
1573: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1574: info->fill_ratio_needed = 0;
1575: info->factor_mallocs = 0;
1576: return(0);
1577: }
1581: PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg)
1582: {
1583: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1587: switch (op) {
1588: case MAT_NEW_NONZERO_LOCATIONS:
1589: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1590: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1591: case MAT_KEEP_NONZERO_PATTERN:
1592: case MAT_NEW_NONZERO_LOCATION_ERR:
1593: MatSetOption(a->A,op,flg);
1594: MatSetOption(a->B,op,flg);
1595: break;
1596: case MAT_ROW_ORIENTED:
1597: a->roworiented = flg;
1599: MatSetOption(a->A,op,flg);
1600: MatSetOption(a->B,op,flg);
1601: break;
1602: case MAT_NEW_DIAGONALS:
1603: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1604: break;
1605: case MAT_IGNORE_OFF_PROC_ENTRIES:
1606: a->donotstash = flg;
1607: break;
1608: case MAT_USE_HASH_TABLE:
1609: a->ht_flag = flg;
1610: break;
1611: case MAT_SYMMETRIC:
1612: case MAT_STRUCTURALLY_SYMMETRIC:
1613: case MAT_HERMITIAN:
1614: case MAT_SYMMETRY_ETERNAL:
1615: MatSetOption(a->A,op,flg);
1616: break;
1617: default:
1618: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"unknown option %d",op);
1619: }
1620: return(0);
1621: }
1625: PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout)
1626: {
1627: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)A->data;
1628: Mat_SeqBAIJ *Aloc;
1629: Mat B;
1631: PetscInt M =A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col;
1632: PetscInt bs=A->rmap->bs,mbs=baij->mbs;
1633: MatScalar *a;
1636: if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1637: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1638: MatCreate(PetscObjectComm((PetscObject)A),&B);
1639: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
1640: MatSetType(B,((PetscObject)A)->type_name);
1641: /* Do not know preallocation information, but must set block size */
1642: MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,NULL,PETSC_DECIDE,NULL);
1643: } else {
1644: B = *matout;
1645: }
1647: /* copy over the A part */
1648: Aloc = (Mat_SeqBAIJ*)baij->A->data;
1649: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1650: PetscMalloc(bs*sizeof(PetscInt),&rvals);
1652: for (i=0; i<mbs; i++) {
1653: rvals[0] = bs*(baij->rstartbs + i);
1654: for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1655: for (j=ai[i]; j<ai[i+1]; j++) {
1656: col = (baij->cstartbs+aj[j])*bs;
1657: for (k=0; k<bs; k++) {
1658: MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);
1660: col++; a += bs;
1661: }
1662: }
1663: }
1664: /* copy over the B part */
1665: Aloc = (Mat_SeqBAIJ*)baij->B->data;
1666: ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1667: for (i=0; i<mbs; i++) {
1668: rvals[0] = bs*(baij->rstartbs + i);
1669: for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1670: for (j=ai[i]; j<ai[i+1]; j++) {
1671: col = baij->garray[aj[j]]*bs;
1672: for (k=0; k<bs; k++) {
1673: MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);
1674: col++;
1675: a += bs;
1676: }
1677: }
1678: }
1679: PetscFree(rvals);
1680: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1681: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
1683: if (reuse == MAT_INITIAL_MATRIX || *matout != A) *matout = B;
1684: else {
1685: MatHeaderMerge(A,B);
1686: }
1687: return(0);
1688: }
1692: PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr)
1693: {
1694: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
1695: Mat a = baij->A,b = baij->B;
1697: PetscInt s1,s2,s3;
1700: MatGetLocalSize(mat,&s2,&s3);
1701: if (rr) {
1702: VecGetLocalSize(rr,&s1);
1703: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1704: /* Overlap communication with computation. */
1705: VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1706: }
1707: if (ll) {
1708: VecGetLocalSize(ll,&s1);
1709: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1710: (*b->ops->diagonalscale)(b,ll,NULL);
1711: }
1712: /* scale the diagonal block */
1713: (*a->ops->diagonalscale)(a,ll,rr);
1715: if (rr) {
1716: /* Do a scatter end and then right scale the off-diagonal block */
1717: VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1718: (*b->ops->diagonalscale)(b,NULL,baij->lvec);
1719: }
1720: return(0);
1721: }
1725: PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1726: {
1727: Mat_MPIBAIJ *l = (Mat_MPIBAIJ*)A->data;
1728: PetscErrorCode ierr;
1729: PetscMPIInt imdex,size = l->size,n,rank = l->rank;
1730: PetscInt i,*owners = A->rmap->range;
1731: PetscInt *nprocs,j,idx,nsends,row;
1732: PetscInt nmax,*svalues,*starts,*owner,nrecvs;
1733: PetscInt *rvalues,tag = ((PetscObject)A)->tag,count,base,slen,*source,lastidx = -1;
1734: PetscInt *lens,*lrows,*values,rstart_bs=A->rmap->rstart;
1735: MPI_Comm comm;
1736: MPI_Request *send_waits,*recv_waits;
1737: MPI_Status recv_status,*send_status;
1738: const PetscScalar *xx;
1739: PetscScalar *bb;
1740: #if defined(PETSC_DEBUG)
1741: PetscBool found = PETSC_FALSE;
1742: #endif
1745: PetscObjectGetComm((PetscObject)A,&comm);
1746: /* first count number of contributors to each processor */
1747: PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
1748: PetscMemzero(nprocs,2*size*sizeof(PetscInt));
1749: PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
1750: j = 0;
1751: for (i=0; i<N; i++) {
1752: if (lastidx > (idx = rows[i])) j = 0;
1753: lastidx = idx;
1754: for (; j<size; j++) {
1755: if (idx >= owners[j] && idx < owners[j+1]) {
1756: nprocs[2*j]++;
1757: nprocs[2*j+1] = 1;
1758: owner[i] = j;
1759: #if defined(PETSC_DEBUG)
1760: found = PETSC_TRUE;
1761: #endif
1762: break;
1763: }
1764: }
1765: #if defined(PETSC_DEBUG)
1766: if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
1767: found = PETSC_FALSE;
1768: #endif
1769: }
1770: nsends = 0; for (i=0; i<size; i++) nsends += nprocs[2*i+1];
1772: if (A->nooffproczerorows) {
1773: if (nsends > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"You called MatSetOption(,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE) but set an off process zero row");
1774: nrecvs = nsends;
1775: nmax = N;
1776: } else {
1777: /* inform other processors of number of messages and max length*/
1778: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
1779: }
1781: /* post receives: */
1782: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
1783: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
1784: for (i=0; i<nrecvs; i++) {
1785: MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
1786: }
1788: /* do sends:
1789: 1) starts[i] gives the starting index in svalues for stuff going to
1790: the ith processor
1791: */
1792: PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
1793: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
1794: PetscMalloc((size+1)*sizeof(PetscInt),&starts);
1795: starts[0] = 0;
1796: for (i=1; i<size; i++) starts[i] = starts[i-1] + nprocs[2*i-2];
1797: for (i=0; i<N; i++) {
1798: svalues[starts[owner[i]]++] = rows[i];
1799: }
1801: starts[0] = 0;
1802: for (i=1; i<size+1; i++) starts[i] = starts[i-1] + nprocs[2*i-2];
1803: count = 0;
1804: for (i=0; i<size; i++) {
1805: if (nprocs[2*i+1]) {
1806: MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
1807: }
1808: }
1809: PetscFree(starts);
1811: base = owners[rank];
1813: /* wait on receives */
1814: PetscMalloc2(nrecvs+1,PetscInt,&lens,nrecvs+1,PetscInt,&source);
1815: count = nrecvs;
1816: slen = 0;
1817: while (count) {
1818: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
1819: /* unpack receives into our local space */
1820: MPI_Get_count(&recv_status,MPIU_INT,&n);
1822: source[imdex] = recv_status.MPI_SOURCE;
1823: lens[imdex] = n;
1824: slen += n;
1825: count--;
1826: }
1827: PetscFree(recv_waits);
1829: /* move the data into the send scatter */
1830: PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
1831: count = 0;
1832: for (i=0; i<nrecvs; i++) {
1833: values = rvalues + i*nmax;
1834: for (j=0; j<lens[i]; j++) {
1835: lrows[count++] = values[j] - base;
1836: }
1837: }
1838: PetscFree(rvalues);
1839: PetscFree2(lens,source);
1840: PetscFree(owner);
1841: PetscFree(nprocs);
1843: /* fix right hand side if needed */
1844: if (x && b) {
1845: VecGetArrayRead(x,&xx);
1846: VecGetArray(b,&bb);
1847: for (i=0; i<slen; i++) {
1848: bb[lrows[i]] = diag*xx[lrows[i]];
1849: }
1850: VecRestoreArrayRead(x,&xx);
1851: VecRestoreArray(b,&bb);
1852: }
1854: /* actually zap the local rows */
1855: /*
1856: Zero the required rows. If the "diagonal block" of the matrix
1857: is square and the user wishes to set the diagonal we use separate
1858: code so that MatSetValues() is not called for each diagonal allocating
1859: new memory, thus calling lots of mallocs and slowing things down.
1861: */
1862: /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
1863: MatZeroRows_SeqBAIJ(l->B,slen,lrows,0.0,0,0);
1864: if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) {
1865: MatZeroRows_SeqBAIJ(l->A,slen,lrows,diag,0,0);
1866: } else if (diag != 0.0) {
1867: MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);
1868: if (((Mat_SeqBAIJ*)l->A->data)->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\
1869: MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
1870: for (i=0; i<slen; i++) {
1871: row = lrows[i] + rstart_bs;
1872: MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
1873: }
1874: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1875: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1876: } else {
1877: MatZeroRows_SeqBAIJ(l->A,slen,lrows,0.0,0,0);
1878: }
1880: PetscFree(lrows);
1882: /* wait on sends */
1883: if (nsends) {
1884: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
1885: MPI_Waitall(nsends,send_waits,send_status);
1886: PetscFree(send_status);
1887: }
1888: PetscFree(send_waits);
1889: PetscFree(svalues);
1890: return(0);
1891: }
1895: PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A)
1896: {
1897: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1901: MatSetUnfactored(a->A);
1902: return(0);
1903: }
1905: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat*);
1909: PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool *flag)
1910: {
1911: Mat_MPIBAIJ *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data;
1912: Mat a,b,c,d;
1913: PetscBool flg;
1917: a = matA->A; b = matA->B;
1918: c = matB->A; d = matB->B;
1920: MatEqual(a,c,&flg);
1921: if (flg) {
1922: MatEqual(b,d,&flg);
1923: }
1924: MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
1925: return(0);
1926: }
1930: PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str)
1931: {
1933: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1934: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data;
1937: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1938: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1939: MatCopy_Basic(A,B,str);
1940: } else {
1941: MatCopy(a->A,b->A,str);
1942: MatCopy(a->B,b->B,str);
1943: }
1944: return(0);
1945: }
1949: PetscErrorCode MatSetUp_MPIBAIJ(Mat A)
1950: {
1954: MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1955: return(0);
1956: }
1960: PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1961: {
1963: Mat_MPIBAIJ *xx=(Mat_MPIBAIJ*)X->data,*yy=(Mat_MPIBAIJ*)Y->data;
1964: PetscBLASInt bnz,one=1;
1965: Mat_SeqBAIJ *x,*y;
1968: if (str == SAME_NONZERO_PATTERN) {
1969: PetscScalar alpha = a;
1970: x = (Mat_SeqBAIJ*)xx->A->data;
1971: y = (Mat_SeqBAIJ*)yy->A->data;
1972: PetscBLASIntCast(x->nz,&bnz);
1973: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
1974: x = (Mat_SeqBAIJ*)xx->B->data;
1975: y = (Mat_SeqBAIJ*)yy->B->data;
1976: PetscBLASIntCast(x->nz,&bnz);
1977: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
1978: } else {
1979: MatAXPY_Basic(Y,a,X,str);
1980: }
1981: return(0);
1982: }
1986: PetscErrorCode MatRealPart_MPIBAIJ(Mat A)
1987: {
1988: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
1992: MatRealPart(a->A);
1993: MatRealPart(a->B);
1994: return(0);
1995: }
1999: PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A)
2000: {
2001: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
2005: MatImaginaryPart(a->A);
2006: MatImaginaryPart(a->B);
2007: return(0);
2008: }
2012: PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
2013: {
2015: IS iscol_local;
2016: PetscInt csize;
2019: ISGetLocalSize(iscol,&csize);
2020: if (call == MAT_REUSE_MATRIX) {
2021: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
2022: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2023: } else {
2024: ISAllGather(iscol,&iscol_local);
2025: }
2026: MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
2027: if (call == MAT_INITIAL_MATRIX) {
2028: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
2029: ISDestroy(&iscol_local);
2030: }
2031: return(0);
2032: }
2033: extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*);
2036: /*
2037: Not great since it makes two copies of the submatrix, first an SeqBAIJ
2038: in local and then by concatenating the local matrices the end result.
2039: Writing it directly would be much like MatGetSubMatrices_MPIBAIJ()
2040: */
2041: PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2042: {
2044: PetscMPIInt rank,size;
2045: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs;
2046: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow;
2047: Mat M,Mreuse;
2048: MatScalar *vwork,*aa;
2049: MPI_Comm comm;
2050: IS isrow_new, iscol_new;
2051: PetscBool idflag,allrows, allcols;
2052: Mat_SeqBAIJ *aij;
2055: PetscObjectGetComm((PetscObject)mat,&comm);
2056: MPI_Comm_rank(comm,&rank);
2057: MPI_Comm_size(comm,&size);
2058: /* The compression and expansion should be avoided. Doesn't point
2059: out errors, might change the indices, hence buggey */
2060: ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);
2061: ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);
2063: /* Check for special case: each processor gets entire matrix columns */
2064: ISIdentity(iscol,&idflag);
2065: ISGetLocalSize(iscol,&ncol);
2066: if (idflag && ncol == mat->cmap->N) allcols = PETSC_TRUE;
2067: else allcols = PETSC_FALSE;
2069: ISIdentity(isrow,&idflag);
2070: ISGetLocalSize(isrow,&nrow);
2071: if (idflag && nrow == mat->rmap->N) allrows = PETSC_TRUE;
2072: else allrows = PETSC_FALSE;
2074: if (call == MAT_REUSE_MATRIX) {
2075: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
2076: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2077: MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);
2078: } else {
2079: MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);
2080: }
2081: ISDestroy(&isrow_new);
2082: ISDestroy(&iscol_new);
2083: /*
2084: m - number of local rows
2085: n - number of columns (same on all processors)
2086: rstart - first row in new global matrix generated
2087: */
2088: MatGetBlockSize(mat,&bs);
2089: MatGetSize(Mreuse,&m,&n);
2090: m = m/bs;
2091: n = n/bs;
2093: if (call == MAT_INITIAL_MATRIX) {
2094: aij = (Mat_SeqBAIJ*)(Mreuse)->data;
2095: ii = aij->i;
2096: jj = aij->j;
2098: /*
2099: Determine the number of non-zeros in the diagonal and off-diagonal
2100: portions of the matrix in order to do correct preallocation
2101: */
2103: /* first get start and end of "diagonal" columns */
2104: if (csize == PETSC_DECIDE) {
2105: ISGetSize(isrow,&mglobal);
2106: if (mglobal == n*bs) { /* square matrix */
2107: nlocal = m;
2108: } else {
2109: nlocal = n/size + ((n % size) > rank);
2110: }
2111: } else {
2112: nlocal = csize/bs;
2113: }
2114: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2115: rstart = rend - nlocal;
2116: if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
2118: /* next, compute all the lengths */
2119: PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
2120: olens = dlens + m;
2121: for (i=0; i<m; i++) {
2122: jend = ii[i+1] - ii[i];
2123: olen = 0;
2124: dlen = 0;
2125: for (j=0; j<jend; j++) {
2126: if (*jj < rstart || *jj >= rend) olen++;
2127: else dlen++;
2128: jj++;
2129: }
2130: olens[i] = olen;
2131: dlens[i] = dlen;
2132: }
2133: MatCreate(comm,&M);
2134: MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);
2135: MatSetType(M,((PetscObject)mat)->type_name);
2136: MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);
2137: PetscFree(dlens);
2138: } else {
2139: PetscInt ml,nl;
2141: M = *newmat;
2142: MatGetLocalSize(M,&ml,&nl);
2143: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2144: MatZeroEntries(M);
2145: /*
2146: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2147: rather than the slower MatSetValues().
2148: */
2149: M->was_assembled = PETSC_TRUE;
2150: M->assembled = PETSC_FALSE;
2151: }
2152: MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);
2153: MatGetOwnershipRange(M,&rstart,&rend);
2154: aij = (Mat_SeqBAIJ*)(Mreuse)->data;
2155: ii = aij->i;
2156: jj = aij->j;
2157: aa = aij->a;
2158: for (i=0; i<m; i++) {
2159: row = rstart/bs + i;
2160: nz = ii[i+1] - ii[i];
2161: cwork = jj; jj += nz;
2162: vwork = aa; aa += nz*bs*bs;
2163: MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2164: }
2166: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2167: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2168: *newmat = M;
2170: /* save submatrix used in processor for next request */
2171: if (call == MAT_INITIAL_MATRIX) {
2172: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2173: PetscObjectDereference((PetscObject)Mreuse);
2174: }
2175: return(0);
2176: }
2180: PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B)
2181: {
2182: MPI_Comm comm,pcomm;
2183: PetscInt first,rlocal_size,clocal_size,nrows;
2184: const PetscInt *rows;
2185: PetscMPIInt size;
2186: IS crowp,growp,irowp,lrowp,lcolp;
2190: PetscObjectGetComm((PetscObject)A,&comm);
2191: /* make a collective version of 'rowp' */
2192: PetscObjectGetComm((PetscObject)rowp,&pcomm);
2193: if (pcomm==comm) {
2194: crowp = rowp;
2195: } else {
2196: ISGetSize(rowp,&nrows);
2197: ISGetIndices(rowp,&rows);
2198: ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);
2199: ISRestoreIndices(rowp,&rows);
2200: }
2201: /* collect the global row permutation and invert it */
2202: ISAllGather(crowp,&growp);
2203: ISSetPermutation(growp);
2204: if (pcomm!=comm) {
2205: ISDestroy(&crowp);
2206: }
2207: ISInvertPermutation(growp,PETSC_DECIDE,&irowp);
2208: ISDestroy(&growp);
2209: /* get the local target indices */
2210: MatGetOwnershipRange(A,&first,NULL);
2211: MatGetLocalSize(A,&rlocal_size,&clocal_size);
2212: ISGetIndices(irowp,&rows);
2213: ISCreateGeneral(MPI_COMM_SELF,rlocal_size,rows+first,PETSC_COPY_VALUES,&lrowp);
2214: ISRestoreIndices(irowp,&rows);
2215: ISDestroy(&irowp);
2216: /* the column permutation is so much easier;
2217: make a local version of 'colp' and invert it */
2218: PetscObjectGetComm((PetscObject)colp,&pcomm);
2219: MPI_Comm_size(pcomm,&size);
2220: if (size==1) {
2221: lcolp = colp;
2222: } else {
2223: ISAllGather(colp,&lcolp);
2224: }
2225: ISSetPermutation(lcolp);
2226: /* now we just get the submatrix */
2227: MatGetSubMatrix_MPIBAIJ_Private(A,lrowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);
2228: if (size>1) {
2229: ISDestroy(&lcolp);
2230: }
2231: /* clean up */
2232: ISDestroy(&lrowp);
2233: return(0);
2234: }
2238: PetscErrorCode MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
2239: {
2240: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data;
2241: Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)baij->B->data;
2244: if (nghosts) *nghosts = B->nbs;
2245: if (ghosts) *ghosts = baij->garray;
2246: return(0);
2247: }
2249: extern PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat);
2253: /*
2254: This routine is almost identical to MatFDColoringCreate_MPIBAIJ()!
2255: */
2256: PetscErrorCode MatFDColoringCreate_MPIBAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
2257: {
2258: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
2259: PetscErrorCode ierr;
2260: PetscMPIInt size,*ncolsonproc,*disp,nn;
2261: PetscInt bs,i,n,nrows,j,k,m,ncols,col;
2262: const PetscInt *is,*rows = 0,*A_ci,*A_cj,*B_ci,*B_cj,*ltog;
2263: PetscInt nis = iscoloring->n,nctot,*cols;
2264: PetscInt *rowhit,M,cstart,cend,colb;
2265: PetscInt *columnsforrow,l;
2266: IS *isa;
2267: PetscBool done,flg;
2268: ISLocalToGlobalMapping map = mat->cmap->bmapping;
2269: PetscInt ctype=c->ctype;
2272: if (!mat->assembled) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Matrix must be assembled first; MatAssemblyBegin/End();");
2273: if (ctype == IS_COLORING_GHOSTED && !map) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMappingBlock");
2275: if (map) {ISLocalToGlobalMappingGetIndices(map,<og);}
2276: else ltog = NULL;
2277: ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);
2278: MatGetBlockSize(mat,&bs);
2280: M = mat->rmap->n/bs;
2281: cstart = mat->cmap->rstart/bs;
2282: cend = mat->cmap->rend/bs;
2283: c->M = mat->rmap->N/bs; /* set the global rows and columns and local rows */
2284: c->N = mat->cmap->N/bs;
2285: c->m = mat->rmap->n/bs;
2286: c->rstart = mat->rmap->rstart/bs;
2288: c->ncolors = nis;
2289: PetscMalloc(nis*sizeof(PetscInt),&c->ncolumns);
2290: PetscMalloc(nis*sizeof(PetscInt*),&c->columns);
2291: PetscMalloc(nis*sizeof(PetscInt),&c->nrows);
2292: PetscMalloc(nis*sizeof(PetscInt*),&c->rows);
2293: PetscMalloc(nis*sizeof(PetscInt*),&c->columnsforrow);
2294: PetscLogObjectMemory(c,5*nis*sizeof(PetscInt));
2296: /* Allow access to data structures of local part of matrix */
2297: if (!baij->colmap) {
2298: MatCreateColmap_MPIBAIJ_Private(mat);
2299: }
2300: MatGetColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);
2301: MatGetColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);
2303: PetscMalloc((M+1)*sizeof(PetscInt),&rowhit);
2304: PetscMalloc((M+1)*sizeof(PetscInt),&columnsforrow);
2306: for (i=0; i<nis; i++) {
2307: ISGetLocalSize(isa[i],&n);
2308: ISGetIndices(isa[i],&is);
2310: c->ncolumns[i] = n;
2311: if (n) {
2312: PetscMalloc(n*sizeof(PetscInt),&c->columns[i]);
2313: PetscLogObjectMemory(c,n*sizeof(PetscInt));
2314: PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));
2315: } else {
2316: c->columns[i] = 0;
2317: }
2319: if (ctype == IS_COLORING_GLOBAL) {
2320: /* Determine the total (parallel) number of columns of this color */
2321: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
2322: PetscMalloc2(size,PetscMPIInt,&ncolsonproc,size,PetscMPIInt,&disp);
2324: PetscMPIIntCast(n,&nn);
2325: MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,PetscObjectComm((PetscObject)mat));
2326: nctot = 0; for (j=0; j<size; j++) nctot += ncolsonproc[j];
2327: if (!nctot) {
2328: PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");
2329: }
2331: disp[0] = 0;
2332: for (j=1; j<size; j++) {
2333: disp[j] = disp[j-1] + ncolsonproc[j-1];
2334: }
2336: /* Get complete list of columns for color on each processor */
2337: PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);
2338: MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,PetscObjectComm((PetscObject)mat));
2339: PetscFree2(ncolsonproc,disp);
2340: } else if (ctype == IS_COLORING_GHOSTED) {
2341: /* Determine local number of columns of this color on this process, including ghost points */
2342: nctot = n;
2343: PetscMalloc((nctot+1)*sizeof(PetscInt),&cols);
2344: PetscMemcpy(cols,is,n*sizeof(PetscInt));
2345: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type");
2347: /*
2348: Mark all rows affect by these columns
2349: */
2350: /* Temporary option to allow for debugging/testing */
2351: flg = PETSC_FALSE;
2352: PetscOptionsGetBool(NULL,"-matfdcoloring_slow",&flg,NULL);
2353: if (!flg) { /*-----------------------------------------------------------------------------*/
2354: /* crude, fast version */
2355: PetscMemzero(rowhit,M*sizeof(PetscInt));
2356: /* loop over columns*/
2357: for (j=0; j<nctot; j++) {
2358: if (ctype == IS_COLORING_GHOSTED) {
2359: col = ltog[cols[j]];
2360: } else {
2361: col = cols[j];
2362: }
2363: if (col >= cstart && col < cend) {
2364: /* column is in diagonal block of matrix */
2365: rows = A_cj + A_ci[col-cstart];
2366: m = A_ci[col-cstart+1] - A_ci[col-cstart];
2367: } else {
2368: #if defined(PETSC_USE_CTABLE)
2369: PetscTableFind(baij->colmap,col+1,&colb);
2370: colb--;
2371: #else
2372: colb = baij->colmap[col] - 1;
2373: #endif
2374: if (colb == -1) {
2375: m = 0;
2376: } else {
2377: colb = colb/bs;
2378: rows = B_cj + B_ci[colb];
2379: m = B_ci[colb+1] - B_ci[colb];
2380: }
2381: }
2382: /* loop over columns marking them in rowhit */
2383: for (k=0; k<m; k++) {
2384: rowhit[*rows++] = col + 1;
2385: }
2386: }
2388: /* count the number of hits */
2389: nrows = 0;
2390: for (j=0; j<M; j++) {
2391: if (rowhit[j]) nrows++;
2392: }
2393: c->nrows[i] = nrows;
2394: PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);
2395: PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);
2396: PetscLogObjectMemory(c,2*(nrows+1)*sizeof(PetscInt));
2397: nrows = 0;
2398: for (j=0; j<M; j++) {
2399: if (rowhit[j]) {
2400: c->rows[i][nrows] = j;
2401: c->columnsforrow[i][nrows] = rowhit[j] - 1;
2402: nrows++;
2403: }
2404: }
2405: } else { /*-------------------------------------------------------------------------------*/
2406: /* slow version, using rowhit as a linked list */
2407: PetscInt currentcol,fm,mfm;
2408: rowhit[M] = M;
2409: nrows = 0;
2410: /* loop over columns*/
2411: for (j=0; j<nctot; j++) {
2412: if (ctype == IS_COLORING_GHOSTED) {
2413: col = ltog[cols[j]];
2414: } else {
2415: col = cols[j];
2416: }
2417: if (col >= cstart && col < cend) {
2418: /* column is in diagonal block of matrix */
2419: rows = A_cj + A_ci[col-cstart];
2420: m = A_ci[col-cstart+1] - A_ci[col-cstart];
2421: } else {
2422: #if defined(PETSC_USE_CTABLE)
2423: PetscTableFind(baij->colmap,col+1,&colb);
2424: colb--;
2425: #else
2426: colb = baij->colmap[col] - 1;
2427: #endif
2428: if (colb == -1) {
2429: m = 0;
2430: } else {
2431: colb = colb/bs;
2432: rows = B_cj + B_ci[colb];
2433: m = B_ci[colb+1] - B_ci[colb];
2434: }
2435: }
2437: /* loop over columns marking them in rowhit */
2438: fm = M; /* fm points to first entry in linked list */
2439: for (k=0; k<m; k++) {
2440: currentcol = *rows++;
2441: /* is it already in the list? */
2442: do {
2443: mfm = fm;
2444: fm = rowhit[fm];
2445: } while (fm < currentcol);
2446: /* not in list so add it */
2447: if (fm != currentcol) {
2448: nrows++;
2449: columnsforrow[currentcol] = col;
2450: /* next three lines insert new entry into linked list */
2451: rowhit[mfm] = currentcol;
2452: rowhit[currentcol] = fm;
2453: fm = currentcol;
2454: /* fm points to present position in list since we know the columns are sorted */
2455: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid coloring of matrix detected");
2456: }
2457: }
2458: c->nrows[i] = nrows;
2459: PetscMalloc((nrows+1)*sizeof(PetscInt),&c->rows[i]);
2460: PetscMalloc((nrows+1)*sizeof(PetscInt),&c->columnsforrow[i]);
2461: PetscLogObjectMemory(c,(nrows+1)*sizeof(PetscInt));
2462: /* now store the linked list of rows into c->rows[i] */
2463: nrows = 0;
2464: fm = rowhit[M];
2465: do {
2466: c->rows[i][nrows] = fm;
2467: c->columnsforrow[i][nrows++] = columnsforrow[fm];
2468: fm = rowhit[fm];
2469: } while (fm < M);
2470: } /* ---------------------------------------------------------------------------------------*/
2471: PetscFree(cols);
2472: }
2474: /* Optimize by adding the vscale, and scaleforrow[][] fields */
2475: /*
2476: vscale will contain the "diagonal" on processor scalings followed by the off processor
2477: */
2478: if (ctype == IS_COLORING_GLOBAL) {
2479: PetscInt *garray;
2480: PetscMalloc(baij->B->cmap->n*sizeof(PetscInt),&garray);
2481: for (i=0; i<baij->B->cmap->n/bs; i++) {
2482: for (j=0; j<bs; j++) {
2483: garray[i*bs+j] = bs*baij->garray[i]+j;
2484: }
2485: }
2486: VecCreateGhost(PetscObjectComm((PetscObject)mat),baij->A->rmap->n,PETSC_DETERMINE,baij->B->cmap->n,garray,&c->vscale);
2487: PetscFree(garray);
2488: PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);
2489: for (k=0; k<c->ncolors; k++) {
2490: PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);
2491: for (l=0; l<c->nrows[k]; l++) {
2492: col = c->columnsforrow[k][l];
2493: if (col >= cstart && col < cend) {
2494: /* column is in diagonal block of matrix */
2495: colb = col - cstart;
2496: } else {
2497: /* column is in "off-processor" part */
2498: #if defined(PETSC_USE_CTABLE)
2499: PetscTableFind(baij->colmap,col+1,&colb);
2500: colb--;
2501: #else
2502: colb = baij->colmap[col] - 1;
2503: #endif
2504: colb = colb/bs;
2505: colb += cend - cstart;
2506: }
2507: c->vscaleforrow[k][l] = colb;
2508: }
2509: }
2510: } else if (ctype == IS_COLORING_GHOSTED) {
2511: /* Get gtol mapping */
2512: PetscInt N = mat->cmap->N,nlocal,*gtol;
2513: PetscMalloc((N+1)*sizeof(PetscInt),>ol);
2514: for (i=0; i<N; i++) gtol[i] = -1;
2515: ISLocalToGlobalMappingGetSize(map,&nlocal);
2516: for (i=0; i<nlocal; i++) gtol[ltog[i]] = i;
2518: c->vscale = 0; /* will be created in MatFDColoringApply() */
2519: PetscMalloc(c->ncolors*sizeof(PetscInt*),&c->vscaleforrow);
2520: for (k=0; k<c->ncolors; k++) {
2521: PetscMalloc((c->nrows[k]+1)*sizeof(PetscInt),&c->vscaleforrow[k]);
2522: for (l=0; l<c->nrows[k]; l++) {
2523: col = c->columnsforrow[k][l]; /* global column index */
2525: c->vscaleforrow[k][l] = gtol[col]; /* local column index */
2526: }
2527: }
2528: PetscFree(gtol);
2529: }
2530: ISColoringRestoreIS(iscoloring,&isa);
2532: PetscFree(rowhit);
2533: PetscFree(columnsforrow);
2534: MatRestoreColumnIJ(baij->A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&done);
2535: MatRestoreColumnIJ(baij->B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&done);
2536: if (map) {ISLocalToGlobalMappingRestoreIndices(map,<og);}
2537: return(0);
2538: }
2542: PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat)
2543: {
2544: Mat B;
2545: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
2546: Mat_SeqBAIJ *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data;
2547: Mat_SeqAIJ *b;
2549: PetscMPIInt size,rank,*recvcounts = 0,*displs = 0;
2550: PetscInt sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs;
2551: PetscInt m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf;
2554: MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
2555: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
2557: /* ----------------------------------------------------------------
2558: Tell every processor the number of nonzeros per row
2559: */
2560: PetscMalloc((A->rmap->N/bs)*sizeof(PetscInt),&lens);
2561: for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) {
2562: lens[i] = ad->i[i-A->rmap->rstart/bs+1] - ad->i[i-A->rmap->rstart/bs] + bd->i[i-A->rmap->rstart/bs+1] - bd->i[i-A->rmap->rstart/bs];
2563: }
2564: sendcount = A->rmap->rend/bs - A->rmap->rstart/bs;
2565: PetscMalloc(2*size*sizeof(PetscMPIInt),&recvcounts);
2566: displs = recvcounts + size;
2567: for (i=0; i<size; i++) {
2568: recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs;
2569: displs[i] = A->rmap->range[i]/bs;
2570: }
2571: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2572: MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2573: #else
2574: MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2575: #endif
2576: /* ---------------------------------------------------------------
2577: Create the sequential matrix of the same type as the local block diagonal
2578: */
2579: MatCreate(PETSC_COMM_SELF,&B);
2580: MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);
2581: MatSetType(B,MATSEQAIJ);
2582: MatSeqAIJSetPreallocation(B,0,lens);
2583: b = (Mat_SeqAIJ*)B->data;
2585: /*--------------------------------------------------------------------
2586: Copy my part of matrix column indices over
2587: */
2588: sendcount = ad->nz + bd->nz;
2589: jsendbuf = b->j + b->i[rstarts[rank]/bs];
2590: a_jsendbuf = ad->j;
2591: b_jsendbuf = bd->j;
2592: n = A->rmap->rend/bs - A->rmap->rstart/bs;
2593: cnt = 0;
2594: for (i=0; i<n; i++) {
2596: /* put in lower diagonal portion */
2597: m = bd->i[i+1] - bd->i[i];
2598: while (m > 0) {
2599: /* is it above diagonal (in bd (compressed) numbering) */
2600: if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break;
2601: jsendbuf[cnt++] = garray[*b_jsendbuf++];
2602: m--;
2603: }
2605: /* put in diagonal portion */
2606: for (j=ad->i[i]; j<ad->i[i+1]; j++) {
2607: jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++;
2608: }
2610: /* put in upper diagonal portion */
2611: while (m-- > 0) {
2612: jsendbuf[cnt++] = garray[*b_jsendbuf++];
2613: }
2614: }
2615: if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt);
2617: /*--------------------------------------------------------------------
2618: Gather all column indices to all processors
2619: */
2620: for (i=0; i<size; i++) {
2621: recvcounts[i] = 0;
2622: for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) {
2623: recvcounts[i] += lens[j];
2624: }
2625: }
2626: displs[0] = 0;
2627: for (i=1; i<size; i++) {
2628: displs[i] = displs[i-1] + recvcounts[i-1];
2629: }
2630: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2631: MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2632: #else
2633: MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2634: #endif
2635: /*--------------------------------------------------------------------
2636: Assemble the matrix into useable form (note numerical values not yet set)
2637: */
2638: /* set the b->ilen (length of each row) values */
2639: PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));
2640: /* set the b->i indices */
2641: b->i[0] = 0;
2642: for (i=1; i<=A->rmap->N/bs; i++) {
2643: b->i[i] = b->i[i-1] + lens[i-1];
2644: }
2645: PetscFree(lens);
2646: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2647: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2648: PetscFree(recvcounts);
2650: if (A->symmetric) {
2651: MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);
2652: } else if (A->hermitian) {
2653: MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);
2654: } else if (A->structurally_symmetric) {
2655: MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);
2656: }
2657: *newmat = B;
2658: return(0);
2659: }
2663: PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
2664: {
2665: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)matin->data;
2667: Vec bb1 = 0;
2670: if (flag == SOR_APPLY_UPPER) {
2671: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2672: return(0);
2673: }
2675: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) {
2676: VecDuplicate(bb,&bb1);
2677: }
2679: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
2680: if (flag & SOR_ZERO_INITIAL_GUESS) {
2681: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2682: its--;
2683: }
2685: while (its--) {
2686: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2687: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2689: /* update rhs: bb1 = bb - B*x */
2690: VecScale(mat->lvec,-1.0);
2691: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
2693: /* local sweep */
2694: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
2695: }
2696: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
2697: if (flag & SOR_ZERO_INITIAL_GUESS) {
2698: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2699: its--;
2700: }
2701: while (its--) {
2702: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2703: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2705: /* update rhs: bb1 = bb - B*x */
2706: VecScale(mat->lvec,-1.0);
2707: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
2709: /* local sweep */
2710: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
2711: }
2712: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
2713: if (flag & SOR_ZERO_INITIAL_GUESS) {
2714: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2715: its--;
2716: }
2717: while (its--) {
2718: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2719: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2721: /* update rhs: bb1 = bb - B*x */
2722: VecScale(mat->lvec,-1.0);
2723: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
2725: /* local sweep */
2726: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
2727: }
2728: } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel version of SOR requested not supported");
2730: VecDestroy(&bb1);
2731: return(0);
2732: }
2734: extern PetscErrorCode MatFDColoringApply_BAIJ(Mat,MatFDColoring,Vec,MatStructure*,void*);
2738: PetscErrorCode MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values)
2739: {
2740: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*) A->data;
2744: MatInvertBlockDiagonal(a->A,values);
2745: return(0);
2746: }
2749: /* -------------------------------------------------------------------*/
2750: static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ,
2751: MatGetRow_MPIBAIJ,
2752: MatRestoreRow_MPIBAIJ,
2753: MatMult_MPIBAIJ,
2754: /* 4*/ MatMultAdd_MPIBAIJ,
2755: MatMultTranspose_MPIBAIJ,
2756: MatMultTransposeAdd_MPIBAIJ,
2757: 0,
2758: 0,
2759: 0,
2760: /*10*/ 0,
2761: 0,
2762: 0,
2763: MatSOR_MPIBAIJ,
2764: MatTranspose_MPIBAIJ,
2765: /*15*/ MatGetInfo_MPIBAIJ,
2766: MatEqual_MPIBAIJ,
2767: MatGetDiagonal_MPIBAIJ,
2768: MatDiagonalScale_MPIBAIJ,
2769: MatNorm_MPIBAIJ,
2770: /*20*/ MatAssemblyBegin_MPIBAIJ,
2771: MatAssemblyEnd_MPIBAIJ,
2772: MatSetOption_MPIBAIJ,
2773: MatZeroEntries_MPIBAIJ,
2774: /*24*/ MatZeroRows_MPIBAIJ,
2775: 0,
2776: 0,
2777: 0,
2778: 0,
2779: /*29*/ MatSetUp_MPIBAIJ,
2780: 0,
2781: 0,
2782: 0,
2783: 0,
2784: /*34*/ MatDuplicate_MPIBAIJ,
2785: 0,
2786: 0,
2787: 0,
2788: 0,
2789: /*39*/ MatAXPY_MPIBAIJ,
2790: MatGetSubMatrices_MPIBAIJ,
2791: MatIncreaseOverlap_MPIBAIJ,
2792: MatGetValues_MPIBAIJ,
2793: MatCopy_MPIBAIJ,
2794: /*44*/ 0,
2795: MatScale_MPIBAIJ,
2796: 0,
2797: 0,
2798: 0,
2799: /*49*/ 0,
2800: 0,
2801: 0,
2802: 0,
2803: 0,
2804: /*54*/ MatFDColoringCreate_MPIBAIJ,
2805: 0,
2806: MatSetUnfactored_MPIBAIJ,
2807: MatPermute_MPIBAIJ,
2808: MatSetValuesBlocked_MPIBAIJ,
2809: /*59*/ MatGetSubMatrix_MPIBAIJ,
2810: MatDestroy_MPIBAIJ,
2811: MatView_MPIBAIJ,
2812: 0,
2813: 0,
2814: /*64*/ 0,
2815: 0,
2816: 0,
2817: 0,
2818: 0,
2819: /*69*/ MatGetRowMaxAbs_MPIBAIJ,
2820: 0,
2821: 0,
2822: 0,
2823: 0,
2824: /*74*/ 0,
2825: MatFDColoringApply_BAIJ,
2826: 0,
2827: 0,
2828: 0,
2829: /*79*/ 0,
2830: 0,
2831: 0,
2832: 0,
2833: MatLoad_MPIBAIJ,
2834: /*84*/ 0,
2835: 0,
2836: 0,
2837: 0,
2838: 0,
2839: /*89*/ 0,
2840: 0,
2841: 0,
2842: 0,
2843: 0,
2844: /*94*/ 0,
2845: 0,
2846: 0,
2847: 0,
2848: 0,
2849: /*99*/ 0,
2850: 0,
2851: 0,
2852: 0,
2853: 0,
2854: /*104*/0,
2855: MatRealPart_MPIBAIJ,
2856: MatImaginaryPart_MPIBAIJ,
2857: 0,
2858: 0,
2859: /*109*/0,
2860: 0,
2861: 0,
2862: 0,
2863: 0,
2864: /*114*/MatGetSeqNonzeroStructure_MPIBAIJ,
2865: 0,
2866: MatGetGhosts_MPIBAIJ,
2867: 0,
2868: 0,
2869: /*119*/0,
2870: 0,
2871: 0,
2872: 0,
2873: 0,
2874: /*124*/0,
2875: 0,
2876: MatInvertBlockDiagonal_MPIBAIJ,
2877: 0,
2878: 0,
2879: /*129*/ 0,
2880: 0,
2881: 0,
2882: 0,
2883: 0,
2884: /*134*/ 0,
2885: 0,
2886: 0,
2887: 0,
2888: 0,
2889: /*139*/ 0,
2890: 0
2891: };
2895: PetscErrorCode MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a)
2896: {
2898: *a = ((Mat_MPIBAIJ*)A->data)->A;
2899: return(0);
2900: }
2902: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*);
2906: PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[])
2907: {
2908: PetscInt m,rstart,cstart,cend;
2909: PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0;
2910: const PetscInt *JJ =0;
2911: PetscScalar *values=0;
2915: PetscLayoutSetBlockSize(B->rmap,bs);
2916: PetscLayoutSetBlockSize(B->cmap,bs);
2917: PetscLayoutSetUp(B->rmap);
2918: PetscLayoutSetUp(B->cmap);
2919: PetscLayoutGetBlockSize(B->rmap,&bs);
2920: m = B->rmap->n/bs;
2921: rstart = B->rmap->rstart/bs;
2922: cstart = B->cmap->rstart/bs;
2923: cend = B->cmap->rend/bs;
2925: if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]);
2926: PetscMalloc2(m,PetscInt,&d_nnz,m,PetscInt,&o_nnz);
2927: for (i=0; i<m; i++) {
2928: nz = ii[i+1] - ii[i];
2929: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz);
2930: nz_max = PetscMax(nz_max,nz);
2931: JJ = jj + ii[i];
2932: for (j=0; j<nz; j++) {
2933: if (*JJ >= cstart) break;
2934: JJ++;
2935: }
2936: d = 0;
2937: for (; j<nz; j++) {
2938: if (*JJ++ >= cend) break;
2939: d++;
2940: }
2941: d_nnz[i] = d;
2942: o_nnz[i] = nz - d;
2943: }
2944: MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);
2945: PetscFree2(d_nnz,o_nnz);
2947: values = (PetscScalar*)V;
2948: if (!values) {
2949: PetscMalloc(bs*bs*nz_max*sizeof(PetscScalar),&values);
2950: PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));
2951: }
2952: for (i=0; i<m; i++) {
2953: PetscInt row = i + rstart;
2954: PetscInt ncols = ii[i+1] - ii[i];
2955: const PetscInt *icols = jj + ii[i];
2956: const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0);
2957: MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);
2958: }
2960: if (!V) { PetscFree(values); }
2961: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2962: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2963: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
2964: return(0);
2965: }
2969: /*@C
2970: MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format
2971: (the default parallel PETSc format).
2973: Collective on MPI_Comm
2975: Input Parameters:
2976: + A - the matrix
2977: . bs - the block size
2978: . i - the indices into j for the start of each local row (starts with zero)
2979: . j - the column indices for each local row (starts with zero) these must be sorted for each row
2980: - v - optional values in the matrix
2982: Level: developer
2984: .keywords: matrix, aij, compressed row, sparse, parallel
2986: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ
2987: @*/
2988: PetscErrorCode MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2989: {
2996: PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));
2997: return(0);
2998: }
3002: PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz)
3003: {
3004: Mat_MPIBAIJ *b;
3006: PetscInt i;
3009: PetscLayoutSetBlockSize(B->rmap,bs);
3010: PetscLayoutSetBlockSize(B->cmap,bs);
3011: PetscLayoutSetUp(B->rmap);
3012: PetscLayoutSetUp(B->cmap);
3013: PetscLayoutGetBlockSize(B->rmap,&bs);
3015: if (d_nnz) {
3016: for (i=0; i<B->rmap->n/bs; i++) {
3017: if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]);
3018: }
3019: }
3020: if (o_nnz) {
3021: for (i=0; i<B->rmap->n/bs; i++) {
3022: if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]);
3023: }
3024: }
3026: b = (Mat_MPIBAIJ*)B->data;
3027: b->bs2 = bs*bs;
3028: b->mbs = B->rmap->n/bs;
3029: b->nbs = B->cmap->n/bs;
3030: b->Mbs = B->rmap->N/bs;
3031: b->Nbs = B->cmap->N/bs;
3033: for (i=0; i<=b->size; i++) {
3034: b->rangebs[i] = B->rmap->range[i]/bs;
3035: }
3036: b->rstartbs = B->rmap->rstart/bs;
3037: b->rendbs = B->rmap->rend/bs;
3038: b->cstartbs = B->cmap->rstart/bs;
3039: b->cendbs = B->cmap->rend/bs;
3041: if (!B->preallocated) {
3042: MatCreate(PETSC_COMM_SELF,&b->A);
3043: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
3044: MatSetType(b->A,MATSEQBAIJ);
3045: PetscLogObjectParent(B,b->A);
3046: MatCreate(PETSC_COMM_SELF,&b->B);
3047: MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
3048: MatSetType(b->B,MATSEQBAIJ);
3049: PetscLogObjectParent(B,b->B);
3050: MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);
3051: }
3053: MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);
3054: MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);
3055: B->preallocated = PETSC_TRUE;
3056: return(0);
3057: }
3059: extern PetscErrorCode MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec);
3060: extern PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal);
3064: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj)
3065: {
3066: Mat_MPIBAIJ *b = (Mat_MPIBAIJ*)B->data;
3068: Mat_SeqBAIJ *d = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data;
3069: PetscInt M = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs;
3070: const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray;
3073: PetscMalloc((M+1)*sizeof(PetscInt),&ii);
3074: ii[0] = 0;
3075: for (i=0; i<M; i++) {
3076: if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]);
3077: if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]);
3078: ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i];
3079: /* remove one from count of matrix has diagonal */
3080: for (j=id[i]; j<id[i+1]; j++) {
3081: if (jd[j] == i) {ii[i+1]--;break;}
3082: }
3083: }
3084: PetscMalloc(ii[M]*sizeof(PetscInt),&jj);
3085: cnt = 0;
3086: for (i=0; i<M; i++) {
3087: for (j=io[i]; j<io[i+1]; j++) {
3088: if (garray[jo[j]] > rstart) break;
3089: jj[cnt++] = garray[jo[j]];
3090: }
3091: for (k=id[i]; k<id[i+1]; k++) {
3092: if (jd[k] != i) {
3093: jj[cnt++] = rstart + jd[k];
3094: }
3095: }
3096: for (; j<io[i+1]; j++) {
3097: jj[cnt++] = garray[jo[j]];
3098: }
3099: }
3100: MatCreateMPIAdj(PetscObjectComm((PetscObject)B),M,B->cmap->N/B->rmap->bs,ii,jj,NULL,adj);
3101: return(0);
3102: }
3104: #include <../src/mat/impls/aij/mpi/mpiaij.h>
3106: PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*);
3110: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat)
3111: {
3113: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
3114: Mat B;
3115: Mat_MPIAIJ *b;
3118: if (!A->assembled) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Matrix must be assembled");
3120: MatCreate(PetscObjectComm((PetscObject)A),&B);
3121: MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
3122: MatSetType(B,MATMPIAIJ);
3123: MatSeqAIJSetPreallocation(B,0,NULL);
3124: MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);
3125: b = (Mat_MPIAIJ*) B->data;
3127: MatDestroy(&b->A);
3128: MatDestroy(&b->B);
3129: MatDisAssemble_MPIBAIJ(A);
3130: MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);
3131: MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);
3132: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3133: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3134: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
3135: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
3136: if (reuse == MAT_REUSE_MATRIX) {
3137: MatHeaderReplace(A,B);
3138: } else {
3139: *newmat = B;
3140: }
3141: return(0);
3142: }
3144: #if defined(PETSC_HAVE_MUMPS)
3145: PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*);
3146: #endif
3148: /*MC
3149: MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices.
3151: Options Database Keys:
3152: + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions()
3153: . -mat_block_size <bs> - set the blocksize used to store the matrix
3154: - -mat_use_hash_table <fact>
3156: Level: beginner
3158: .seealso: MatCreateMPIBAIJ
3159: M*/
3161: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*);
3165: PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B)
3166: {
3167: Mat_MPIBAIJ *b;
3169: PetscBool flg;
3172: PetscNewLog(B,Mat_MPIBAIJ,&b);
3173: B->data = (void*)b;
3175: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3176: B->assembled = PETSC_FALSE;
3178: B->insertmode = NOT_SET_VALUES;
3179: MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
3180: MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);
3182: /* build local table of row and column ownerships */
3183: PetscMalloc((b->size+1)*sizeof(PetscInt),&b->rangebs);
3185: /* build cache for off array entries formed */
3186: MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);
3188: b->donotstash = PETSC_FALSE;
3189: b->colmap = NULL;
3190: b->garray = NULL;
3191: b->roworiented = PETSC_TRUE;
3193: /* stuff used in block assembly */
3194: b->barray = 0;
3196: /* stuff used for matrix vector multiply */
3197: b->lvec = 0;
3198: b->Mvctx = 0;
3200: /* stuff for MatGetRow() */
3201: b->rowindices = 0;
3202: b->rowvalues = 0;
3203: b->getrowactive = PETSC_FALSE;
3205: /* hash table stuff */
3206: b->ht = 0;
3207: b->hd = 0;
3208: b->ht_size = 0;
3209: b->ht_flag = PETSC_FALSE;
3210: b->ht_fact = 0;
3211: b->ht_total_ct = 0;
3212: b->ht_insert_ct = 0;
3214: /* stuff for MatGetSubMatrices_MPIBAIJ_local() */
3215: b->ijonly = PETSC_FALSE;
3217: PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPIBAIJ matrix 1","Mat");
3218: PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,NULL);
3219: if (flg) {
3220: PetscReal fact = 1.39;
3221: MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);
3222: PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);
3223: if (fact <= 1.0) fact = 1.39;
3224: MatMPIBAIJSetHashTableFactor(B,fact);
3225: PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);
3226: }
3227: PetscOptionsEnd();
3229: #if defined(PETSC_HAVE_MUMPS)
3230: PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_baij_mumps);
3231: #endif
3232: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiadj_C",MatConvert_MPIBAIJ_MPIAdj);
3233: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiaij_C",MatConvert_MPIBAIJ_MPIAIJ);
3234: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C",MatConvert_MPIBAIJ_MPISBAIJ);
3235: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIBAIJ);
3236: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIBAIJ);
3237: PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIBAIJ);
3238: PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJ);
3239: PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",MatMPIBAIJSetPreallocationCSR_MPIBAIJ);
3240: PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIBAIJ);
3241: PetscObjectComposeFunction((PetscObject)B,"MatSetHashTableFactor_C",MatSetHashTableFactor_MPIBAIJ);
3242: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C",MatConvert_MPIBAIJ_MPIBSTRM);
3243: PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);
3244: return(0);
3245: }
3247: /*MC
3248: MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices.
3250: This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator,
3251: and MATMPIBAIJ otherwise.
3253: Options Database Keys:
3254: . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions()
3256: Level: beginner
3258: .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3259: M*/
3263: /*@C
3264: MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format
3265: (block compressed row). For good matrix assembly performance
3266: the user should preallocate the matrix storage by setting the parameters
3267: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
3268: performance can be increased by more than a factor of 50.
3270: Collective on Mat
3272: Input Parameters:
3273: + A - the matrix
3274: . bs - size of blockk
3275: . d_nz - number of block nonzeros per block row in diagonal portion of local
3276: submatrix (same for all local rows)
3277: . d_nnz - array containing the number of block nonzeros in the various block rows
3278: of the in diagonal portion of the local (possibly different for each block
3279: row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry and
3280: set it even if it is zero.
3281: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
3282: submatrix (same for all local rows).
3283: - o_nnz - array containing the number of nonzeros in the various block rows of the
3284: off-diagonal portion of the local submatrix (possibly different for
3285: each block row) or NULL.
3287: If the *_nnz parameter is given then the *_nz parameter is ignored
3289: Options Database Keys:
3290: + -mat_block_size - size of the blocks to use
3291: - -mat_use_hash_table <fact>
3293: Notes:
3294: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
3295: than it must be used on all processors that share the object for that argument.
3297: Storage Information:
3298: For a square global matrix we define each processor's diagonal portion
3299: to be its local rows and the corresponding columns (a square submatrix);
3300: each processor's off-diagonal portion encompasses the remainder of the
3301: local matrix (a rectangular submatrix).
3303: The user can specify preallocated storage for the diagonal part of
3304: the local submatrix with either d_nz or d_nnz (not both). Set
3305: d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3306: memory allocation. Likewise, specify preallocated storage for the
3307: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
3309: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3310: the figure below we depict these three local rows and all columns (0-11).
3312: .vb
3313: 0 1 2 3 4 5 6 7 8 9 10 11
3314: --------------------------
3315: row 3 |o o o d d d o o o o o o
3316: row 4 |o o o d d d o o o o o o
3317: row 5 |o o o d d d o o o o o o
3318: --------------------------
3319: .ve
3321: Thus, any entries in the d locations are stored in the d (diagonal)
3322: submatrix, and any entries in the o locations are stored in the
3323: o (off-diagonal) submatrix. Note that the d and the o submatrices are
3324: stored simply in the MATSEQBAIJ format for compressed row storage.
3326: Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3327: and o_nz should indicate the number of block nonzeros per row in the o matrix.
3328: In general, for PDE problems in which most nonzeros are near the diagonal,
3329: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
3330: or you will get TERRIBLE performance; see the users' manual chapter on
3331: matrices.
3333: You can call MatGetInfo() to get information on how effective the preallocation was;
3334: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3335: You can also run with the option -info and look for messages with the string
3336: malloc in them to see if additional memory allocation was needed.
3338: Level: intermediate
3340: .keywords: matrix, block, aij, compressed row, sparse, parallel
3342: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership()
3343: @*/
3344: PetscErrorCode MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3345: {
3352: PetscTryMethod(B,"MatMPIBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));
3353: return(0);
3354: }
3358: /*@C
3359: MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format
3360: (block compressed row). For good matrix assembly performance
3361: the user should preallocate the matrix storage by setting the parameters
3362: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
3363: performance can be increased by more than a factor of 50.
3365: Collective on MPI_Comm
3367: Input Parameters:
3368: + comm - MPI communicator
3369: . bs - size of blockk
3370: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
3371: This value should be the same as the local size used in creating the
3372: y vector for the matrix-vector product y = Ax.
3373: . n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
3374: This value should be the same as the local size used in creating the
3375: x vector for the matrix-vector product y = Ax.
3376: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3377: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3378: . d_nz - number of nonzero blocks per block row in diagonal portion of local
3379: submatrix (same for all local rows)
3380: . d_nnz - array containing the number of nonzero blocks in the various block rows
3381: of the in diagonal portion of the local (possibly different for each block
3382: row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry
3383: and set it even if it is zero.
3384: . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local
3385: submatrix (same for all local rows).
3386: - o_nnz - array containing the number of nonzero blocks in the various block rows of the
3387: off-diagonal portion of the local submatrix (possibly different for
3388: each block row) or NULL.
3390: Output Parameter:
3391: . A - the matrix
3393: Options Database Keys:
3394: + -mat_block_size - size of the blocks to use
3395: - -mat_use_hash_table <fact>
3397: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3398: MatXXXXSetPreallocation() paradgm instead of this routine directly.
3399: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3401: Notes:
3402: If the *_nnz parameter is given then the *_nz parameter is ignored
3404: A nonzero block is any block that as 1 or more nonzeros in it
3406: The user MUST specify either the local or global matrix dimensions
3407: (possibly both).
3409: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor
3410: than it must be used on all processors that share the object for that argument.
3412: Storage Information:
3413: For a square global matrix we define each processor's diagonal portion
3414: to be its local rows and the corresponding columns (a square submatrix);
3415: each processor's off-diagonal portion encompasses the remainder of the
3416: local matrix (a rectangular submatrix).
3418: The user can specify preallocated storage for the diagonal part of
3419: the local submatrix with either d_nz or d_nnz (not both). Set
3420: d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3421: memory allocation. Likewise, specify preallocated storage for the
3422: off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
3424: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3425: the figure below we depict these three local rows and all columns (0-11).
3427: .vb
3428: 0 1 2 3 4 5 6 7 8 9 10 11
3429: --------------------------
3430: row 3 |o o o d d d o o o o o o
3431: row 4 |o o o d d d o o o o o o
3432: row 5 |o o o d d d o o o o o o
3433: --------------------------
3434: .ve
3436: Thus, any entries in the d locations are stored in the d (diagonal)
3437: submatrix, and any entries in the o locations are stored in the
3438: o (off-diagonal) submatrix. Note that the d and the o submatrices are
3439: stored simply in the MATSEQBAIJ format for compressed row storage.
3441: Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3442: and o_nz should indicate the number of block nonzeros per row in the o matrix.
3443: In general, for PDE problems in which most nonzeros are near the diagonal,
3444: one expects d_nz >> o_nz. For large problems you MUST preallocate memory
3445: or you will get TERRIBLE performance; see the users' manual chapter on
3446: matrices.
3448: Level: intermediate
3450: .keywords: matrix, block, aij, compressed row, sparse, parallel
3452: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3453: @*/
3454: PetscErrorCode MatCreateBAIJ(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)
3455: {
3457: PetscMPIInt size;
3460: MatCreate(comm,A);
3461: MatSetSizes(*A,m,n,M,N);
3462: MPI_Comm_size(comm,&size);
3463: if (size > 1) {
3464: MatSetType(*A,MATMPIBAIJ);
3465: MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
3466: } else {
3467: MatSetType(*A,MATSEQBAIJ);
3468: MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
3469: }
3470: return(0);
3471: }
3475: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3476: {
3477: Mat mat;
3478: Mat_MPIBAIJ *a,*oldmat = (Mat_MPIBAIJ*)matin->data;
3480: PetscInt len=0;
3483: *newmat = 0;
3484: MatCreate(PetscObjectComm((PetscObject)matin),&mat);
3485: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3486: MatSetType(mat,((PetscObject)matin)->type_name);
3487: PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
3489: mat->factortype = matin->factortype;
3490: mat->preallocated = PETSC_TRUE;
3491: mat->assembled = PETSC_TRUE;
3492: mat->insertmode = NOT_SET_VALUES;
3494: a = (Mat_MPIBAIJ*)mat->data;
3495: mat->rmap->bs = matin->rmap->bs;
3496: a->bs2 = oldmat->bs2;
3497: a->mbs = oldmat->mbs;
3498: a->nbs = oldmat->nbs;
3499: a->Mbs = oldmat->Mbs;
3500: a->Nbs = oldmat->Nbs;
3502: PetscLayoutReference(matin->rmap,&mat->rmap);
3503: PetscLayoutReference(matin->cmap,&mat->cmap);
3505: a->size = oldmat->size;
3506: a->rank = oldmat->rank;
3507: a->donotstash = oldmat->donotstash;
3508: a->roworiented = oldmat->roworiented;
3509: a->rowindices = 0;
3510: a->rowvalues = 0;
3511: a->getrowactive = PETSC_FALSE;
3512: a->barray = 0;
3513: a->rstartbs = oldmat->rstartbs;
3514: a->rendbs = oldmat->rendbs;
3515: a->cstartbs = oldmat->cstartbs;
3516: a->cendbs = oldmat->cendbs;
3518: /* hash table stuff */
3519: a->ht = 0;
3520: a->hd = 0;
3521: a->ht_size = 0;
3522: a->ht_flag = oldmat->ht_flag;
3523: a->ht_fact = oldmat->ht_fact;
3524: a->ht_total_ct = 0;
3525: a->ht_insert_ct = 0;
3527: PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));
3528: if (oldmat->colmap) {
3529: #if defined(PETSC_USE_CTABLE)
3530: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3531: #else
3532: PetscMalloc((a->Nbs)*sizeof(PetscInt),&a->colmap);
3533: PetscLogObjectMemory(mat,(a->Nbs)*sizeof(PetscInt));
3534: PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));
3535: #endif
3536: } else a->colmap = 0;
3538: if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
3539: PetscMalloc(len*sizeof(PetscInt),&a->garray);
3540: PetscLogObjectMemory(mat,len*sizeof(PetscInt));
3541: PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));
3542: } else a->garray = 0;
3544: MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);
3545: VecDuplicate(oldmat->lvec,&a->lvec);
3546: PetscLogObjectParent(mat,a->lvec);
3547: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3548: PetscLogObjectParent(mat,a->Mvctx);
3550: MatDuplicate(oldmat->A,cpvalues,&a->A);
3551: PetscLogObjectParent(mat,a->A);
3552: MatDuplicate(oldmat->B,cpvalues,&a->B);
3553: PetscLogObjectParent(mat,a->B);
3554: PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3555: *newmat = mat;
3556: return(0);
3557: }
3561: PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer)
3562: {
3564: int fd;
3565: PetscInt i,nz,j,rstart,rend;
3566: PetscScalar *vals,*buf;
3567: MPI_Comm comm;
3568: MPI_Status status;
3569: PetscMPIInt rank,size,maxnz;
3570: PetscInt header[4],*rowlengths = 0,M,N,m,*rowners,*cols;
3571: PetscInt *locrowlens = NULL,*procsnz = NULL,*browners = NULL;
3572: PetscInt jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax;
3573: PetscMPIInt tag = ((PetscObject)viewer)->tag;
3574: PetscInt *dlens = NULL,*odlens = NULL,*mask = NULL,*masked1 = NULL,*masked2 = NULL,rowcount,odcount;
3575: PetscInt dcount,kmax,k,nzcount,tmp,mend,sizesset=1,grows,gcols;
3578: PetscObjectGetComm((PetscObject)viewer,&comm);
3579: PetscOptionsBegin(comm,NULL,"Options for loading MPIBAIJ matrix 2","Mat");
3580: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
3581: PetscOptionsEnd();
3583: MPI_Comm_size(comm,&size);
3584: MPI_Comm_rank(comm,&rank);
3585: if (!rank) {
3586: PetscViewerBinaryGetDescriptor(viewer,&fd);
3587: PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
3588: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3589: }
3591: if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0;
3593: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3594: M = header[1]; N = header[2];
3596: /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */
3597: if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M;
3598: if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N;
3600: /* If global sizes are set, check if they are consistent with that given in the file */
3601: if (sizesset) {
3602: MatGetSize(newmat,&grows,&gcols);
3603: }
3604: if (sizesset && newmat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows);
3605: if (sizesset && newmat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols);
3607: if (M != N) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Can only do square matrices");
3609: /*
3610: This code adds extra rows to make sure the number of rows is
3611: divisible by the blocksize
3612: */
3613: Mbs = M/bs;
3614: extra_rows = bs - M + bs*Mbs;
3615: if (extra_rows == bs) extra_rows = 0;
3616: else Mbs++;
3617: if (extra_rows && !rank) {
3618: PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");
3619: }
3621: /* determine ownership of all rows */
3622: if (newmat->rmap->n < 0) { /* PETSC_DECIDE */
3623: mbs = Mbs/size + ((Mbs % size) > rank);
3624: m = mbs*bs;
3625: } else { /* User set */
3626: m = newmat->rmap->n;
3627: mbs = m/bs;
3628: }
3629: PetscMalloc2(size+1,PetscInt,&rowners,size+1,PetscInt,&browners);
3630: MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
3632: /* process 0 needs enough room for process with most rows */
3633: if (!rank) {
3634: mmax = rowners[1];
3635: for (i=2; i<=size; i++) {
3636: mmax = PetscMax(mmax,rowners[i]);
3637: }
3638: mmax*=bs;
3639: } else mmax = -1; /* unused, but compiler warns anyway */
3641: rowners[0] = 0;
3642: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
3643: for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
3644: rstart = rowners[rank];
3645: rend = rowners[rank+1];
3647: /* distribute row lengths to all processors */
3648: PetscMalloc(m*sizeof(PetscInt),&locrowlens);
3649: if (!rank) {
3650: mend = m;
3651: if (size == 1) mend = mend - extra_rows;
3652: PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);
3653: for (j=mend; j<m; j++) locrowlens[j] = 1;
3654: PetscMalloc(mmax*sizeof(PetscInt),&rowlengths);
3655: PetscMalloc(size*sizeof(PetscInt),&procsnz);
3656: PetscMemzero(procsnz,size*sizeof(PetscInt));
3657: for (j=0; j<m; j++) {
3658: procsnz[0] += locrowlens[j];
3659: }
3660: for (i=1; i<size; i++) {
3661: mend = browners[i+1] - browners[i];
3662: if (i == size-1) mend = mend - extra_rows;
3663: PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);
3664: for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1;
3665: /* calculate the number of nonzeros on each processor */
3666: for (j=0; j<browners[i+1]-browners[i]; j++) {
3667: procsnz[i] += rowlengths[j];
3668: }
3669: MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);
3670: }
3671: PetscFree(rowlengths);
3672: } else {
3673: MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);
3674: }
3676: if (!rank) {
3677: /* determine max buffer needed and allocate it */
3678: maxnz = procsnz[0];
3679: for (i=1; i<size; i++) {
3680: maxnz = PetscMax(maxnz,procsnz[i]);
3681: }
3682: PetscMalloc(maxnz*sizeof(PetscInt),&cols);
3684: /* read in my part of the matrix column indices */
3685: nz = procsnz[0];
3686: PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);
3687: mycols = ibuf;
3688: if (size == 1) nz -= extra_rows;
3689: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3690: if (size == 1) {
3691: for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i;
3692: }
3694: /* read in every ones (except the last) and ship off */
3695: for (i=1; i<size-1; i++) {
3696: nz = procsnz[i];
3697: PetscBinaryRead(fd,cols,nz,PETSC_INT);
3698: MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
3699: }
3700: /* read in the stuff for the last proc */
3701: if (size != 1) {
3702: nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */
3703: PetscBinaryRead(fd,cols,nz,PETSC_INT);
3704: for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
3705: MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);
3706: }
3707: PetscFree(cols);
3708: } else {
3709: /* determine buffer space needed for message */
3710: nz = 0;
3711: for (i=0; i<m; i++) {
3712: nz += locrowlens[i];
3713: }
3714: PetscMalloc((nz+1)*sizeof(PetscInt),&ibuf);
3715: mycols = ibuf;
3716: /* receive message of column indices*/
3717: MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
3718: MPI_Get_count(&status,MPIU_INT,&maxnz);
3719: if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
3720: }
3722: /* loop over local rows, determining number of off diagonal entries */
3723: PetscMalloc2(rend-rstart,PetscInt,&dlens,rend-rstart,PetscInt,&odlens);
3724: PetscMalloc3(Mbs,PetscInt,&mask,Mbs,PetscInt,&masked1,Mbs,PetscInt,&masked2);
3725: PetscMemzero(mask,Mbs*sizeof(PetscInt));
3726: PetscMemzero(masked1,Mbs*sizeof(PetscInt));
3727: PetscMemzero(masked2,Mbs*sizeof(PetscInt));
3728: rowcount = 0; nzcount = 0;
3729: for (i=0; i<mbs; i++) {
3730: dcount = 0;
3731: odcount = 0;
3732: for (j=0; j<bs; j++) {
3733: kmax = locrowlens[rowcount];
3734: for (k=0; k<kmax; k++) {
3735: tmp = mycols[nzcount++]/bs;
3736: if (!mask[tmp]) {
3737: mask[tmp] = 1;
3738: if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp;
3739: else masked1[dcount++] = tmp;
3740: }
3741: }
3742: rowcount++;
3743: }
3745: dlens[i] = dcount;
3746: odlens[i] = odcount;
3748: /* zero out the mask elements we set */
3749: for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
3750: for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
3751: }
3754: if (!sizesset) {
3755: MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);
3756: }
3757: MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);
3759: if (!rank) {
3760: PetscMalloc((maxnz+1)*sizeof(PetscScalar),&buf);
3761: /* read in my part of the matrix numerical values */
3762: nz = procsnz[0];
3763: vals = buf;
3764: mycols = ibuf;
3765: if (size == 1) nz -= extra_rows;
3766: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3767: if (size == 1) {
3768: for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0;
3769: }
3771: /* insert into matrix */
3772: jj = rstart*bs;
3773: for (i=0; i<m; i++) {
3774: MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3775: mycols += locrowlens[i];
3776: vals += locrowlens[i];
3777: jj++;
3778: }
3779: /* read in other processors (except the last one) and ship out */
3780: for (i=1; i<size-1; i++) {
3781: nz = procsnz[i];
3782: vals = buf;
3783: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3784: MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);
3785: }
3786: /* the last proc */
3787: if (size != 1) {
3788: nz = procsnz[i] - extra_rows;
3789: vals = buf;
3790: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3791: for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
3792: MPIULong_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);
3793: }
3794: PetscFree(procsnz);
3795: } else {
3796: /* receive numeric values */
3797: PetscMalloc((nz+1)*sizeof(PetscScalar),&buf);
3799: /* receive message of values*/
3800: vals = buf;
3801: mycols = ibuf;
3802: MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm);
3804: /* insert into matrix */
3805: jj = rstart*bs;
3806: for (i=0; i<m; i++) {
3807: MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3808: mycols += locrowlens[i];
3809: vals += locrowlens[i];
3810: jj++;
3811: }
3812: }
3813: PetscFree(locrowlens);
3814: PetscFree(buf);
3815: PetscFree(ibuf);
3816: PetscFree2(rowners,browners);
3817: PetscFree2(dlens,odlens);
3818: PetscFree3(mask,masked1,masked2);
3819: MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);
3820: MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);
3821: return(0);
3822: }
3826: /*@
3827: MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.
3829: Input Parameters:
3830: . mat - the matrix
3831: . fact - factor
3833: Not Collective, each process can use a different factor
3835: Level: advanced
3837: Notes:
3838: This can also be set by the command line option: -mat_use_hash_table <fact>
3840: .keywords: matrix, hashtable, factor, HT
3842: .seealso: MatSetOption()
3843: @*/
3844: PetscErrorCode MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact)
3845: {
3849: PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));
3850: return(0);
3851: }
3855: PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact)
3856: {
3857: Mat_MPIBAIJ *baij;
3860: baij = (Mat_MPIBAIJ*)mat->data;
3861: baij->ht_fact = fact;
3862: return(0);
3863: }
3867: PetscErrorCode MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
3868: {
3869: Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;
3872: *Ad = a->A;
3873: *Ao = a->B;
3874: *colmap = a->garray;
3875: return(0);
3876: }
3878: /*
3879: Special version for direct calls from Fortran (to eliminate two function call overheads
3880: */
3881: #if defined(PETSC_HAVE_FORTRAN_CAPS)
3882: #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED
3883: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
3884: #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked
3885: #endif
3889: /*@C
3890: MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked()
3892: Collective on Mat
3894: Input Parameters:
3895: + mat - the matrix
3896: . min - number of input rows
3897: . im - input rows
3898: . nin - number of input columns
3899: . in - input columns
3900: . v - numerical values input
3901: - addvin - INSERT_VALUES or ADD_VALUES
3903: Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse.
3905: Level: advanced
3907: .seealso: MatSetValuesBlocked()
3908: @*/
3909: PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin)
3910: {
3911: /* convert input arguments to C version */
3912: Mat mat = *matin;
3913: PetscInt m = *min, n = *nin;
3914: InsertMode addv = *addvin;
3916: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
3917: const MatScalar *value;
3918: MatScalar *barray = baij->barray;
3919: PetscBool roworiented = baij->roworiented;
3920: PetscErrorCode ierr;
3921: PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs;
3922: PetscInt rend=baij->rendbs,cstart=baij->cstartbs,stepval;
3923: PetscInt cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;
3926: /* tasks normally handled by MatSetValuesBlocked() */
3927: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
3928: #if defined(PETSC_USE_DEBUG)
3929: else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
3930: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3931: #endif
3932: if (mat->assembled) {
3933: mat->was_assembled = PETSC_TRUE;
3934: mat->assembled = PETSC_FALSE;
3935: }
3936: PetscLogEventBegin(MAT_SetValues,mat,0,0,0);
3939: if (!barray) {
3940: PetscMalloc(bs2*sizeof(MatScalar),&barray);
3941: baij->barray = barray;
3942: }
3944: if (roworiented) stepval = (n-1)*bs;
3945: else stepval = (m-1)*bs;
3947: for (i=0; i<m; i++) {
3948: if (im[i] < 0) continue;
3949: #if defined(PETSC_USE_DEBUG)
3950: if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
3951: #endif
3952: if (im[i] >= rstart && im[i] < rend) {
3953: row = im[i] - rstart;
3954: for (j=0; j<n; j++) {
3955: /* If NumCol = 1 then a copy is not required */
3956: if ((roworiented) && (n == 1)) {
3957: barray = (MatScalar*)v + i*bs2;
3958: } else if ((!roworiented) && (m == 1)) {
3959: barray = (MatScalar*)v + j*bs2;
3960: } else { /* Here a copy is required */
3961: if (roworiented) {
3962: value = v + i*(stepval+bs)*bs + j*bs;
3963: } else {
3964: value = v + j*(stepval+bs)*bs + i*bs;
3965: }
3966: for (ii=0; ii<bs; ii++,value+=stepval) {
3967: for (jj=0; jj<bs; jj++) {
3968: *barray++ = *value++;
3969: }
3970: }
3971: barray -=bs2;
3972: }
3974: if (in[j] >= cstart && in[j] < cend) {
3975: col = in[j] - cstart;
3976: MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
3977: } else if (in[j] < 0) continue;
3978: #if defined(PETSC_USE_DEBUG)
3979: else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
3980: #endif
3981: else {
3982: if (mat->was_assembled) {
3983: if (!baij->colmap) {
3984: MatCreateColmap_MPIBAIJ_Private(mat);
3985: }
3987: #if defined(PETSC_USE_DEBUG)
3988: #if defined(PETSC_USE_CTABLE)
3989: { PetscInt data;
3990: PetscTableFind(baij->colmap,in[j]+1,&data);
3991: if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3992: }
3993: #else
3994: if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3995: #endif
3996: #endif
3997: #if defined(PETSC_USE_CTABLE)
3998: PetscTableFind(baij->colmap,in[j]+1,&col);
3999: col = (col - 1)/bs;
4000: #else
4001: col = (baij->colmap[in[j]] - 1)/bs;
4002: #endif
4003: if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
4004: MatDisAssemble_MPIBAIJ(mat);
4005: col = in[j];
4006: }
4007: } else col = in[j];
4008: MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
4009: }
4010: }
4011: } else {
4012: if (!baij->donotstash) {
4013: if (roworiented) {
4014: MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
4015: } else {
4016: MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
4017: }
4018: }
4019: }
4020: }
4022: /* task normally handled by MatSetValuesBlocked() */
4023: PetscLogEventEnd(MAT_SetValues,mat,0,0,0);
4024: return(0);
4025: }
4029: /*@
4030: MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard
4031: CSR format the local rows.
4033: Collective on MPI_Comm
4035: Input Parameters:
4036: + comm - MPI communicator
4037: . bs - the block size, only a block size of 1 is supported
4038: . m - number of local rows (Cannot be PETSC_DECIDE)
4039: . n - This value should be the same as the local size used in creating the
4040: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4041: calculated if N is given) For square matrices n is almost always m.
4042: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4043: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4044: . i - row indices
4045: . j - column indices
4046: - a - matrix values
4048: Output Parameter:
4049: . mat - the matrix
4051: Level: intermediate
4053: Notes:
4054: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4055: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4056: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4058: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4060: .keywords: matrix, aij, compressed row, sparse, parallel
4062: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4063: MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
4064: @*/
4065: PetscErrorCode MatCreateMPIBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4066: {
4070: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4071: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4072: MatCreate(comm,mat);
4073: MatSetSizes(*mat,m,n,M,N);
4074: MatSetType(*mat,MATMPISBAIJ);
4075: MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);
4076: return(0);
4077: }