Actual source code: matmatmult.c
1: /*$Id: matmatmult.c,v 1.15 2001/09/07 20:04:44 buschelm Exp $*/
2: /*
3: Defines matrix-matrix product routines for pairs of SeqAIJ matrices
4: C = A * B
5: C = P * A * P^T
6: */
8: #include src/mat/impls/aij/seq/aij.h
9: #include src/mat/utils/freespace.h
11: static int logkey_matmatmult = 0;
12: static int logkey_matmatmult_symbolic = 0;
13: static int logkey_matmatmult_numeric = 0;
15: static int logkey_matapplypapt = 0;
16: static int logkey_matapplypapt_symbolic = 0;
17: static int logkey_matapplypapt_numeric = 0;
19: /*
20: MatMatMult_Symbolic_SeqAIJ_SeqAIJ - Forms the symbolic product of two SeqAIJ matrices
21: C = A * B;
23: Note: C is assumed to be uncreated.
24: If this is not the case, Destroy C before calling this routine.
25: */
26: #ifdef USE_INTSORT
27: /*
28: This roution is modified by the one below for better performance.
29: The changes are:
30: -- PetscSortInt() is replace by a linked list
31: -- malloc larger Initial FreeSpace
32: */
33: #undef __FUNCT__
35: int MatMatMult_Symbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat *C)
36: {
37: int ierr;
38: FreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
39: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
40: int aishift=a->indexshift,bishift=b->indexshift;
41: int *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj;
42: int *ci,*cj,*denserow,*sparserow;
43: int an=A->N,am=A->M,bn=B->N,bm=B->M;
44: int i,j,k,anzi,brow,bnzj,cnzi;
45: MatScalar *ca;
48: /* some error checking which could be moved into interface layer */
49: if (aishift || bishift) SETERRQ(PETSC_ERR_SUP,"Shifted matrix indices are not supported.");
50: if (an!=bm) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",an,bm);
51:
52: /* Set up timers */
53: if (!logkey_matmatmult_symbolic) {
54: PetscLogEventRegister(&logkey_matmatmult_symbolic,"MatMatMult_Symbolic",MAT_COOKIE);
55: }
56: PetscLogEventBegin(logkey_matmatmult_symbolic,A,B,0,0);
58: /* Set up */
59: /* Allocate ci array, arrays for fill computation and */
60: /* free space for accumulating nonzero column info */
61: PetscMalloc(((am+1)+1)*sizeof(int),&ci);
62: ci[0] = 0;
64: PetscMalloc((2*bn+1)*sizeof(int),&denserow);
65: PetscMemzero(denserow,(2*bn+1)*sizeof(int));
66: sparserow = denserow + bn;
68: /* Initial FreeSpace size is nnz(B)=bi[bm] */
69: ierr = GetMoreSpace(bi[bm],&free_space);
70: current_space = free_space;
72: /* Determine symbolic info for each row of the product: */
73: for (i=0;i<am;i++) {
74: anzi = ai[i+1] - ai[i];
75: cnzi = 0;
76: for (j=0;j<anzi;j++) {
77: brow = *aj++;
78: bnzj = bi[brow+1] - bi[brow];
79: bjj = bj + bi[brow];
80: for (k=0;k<bnzj;k++) {
81: /* If column is not marked, mark it in compressed and uncompressed locations. */
82: /* For simplicity, leave uncompressed row unsorted until finished with row, */
83: /* and increment nonzero count for this row. */
84: if (!denserow[bjj[k]]) {
85: denserow[bjj[k]] = -1;
86: sparserow[cnzi++] = bjj[k];
87: }
88: }
89: }
91: /* sort sparserow */
92: PetscSortInt(cnzi,sparserow);
94: /* If free space is not available, make more free space */
95: /* Double the amount of total space in the list */
96: if (current_space->local_remaining<cnzi) {
97: GetMoreSpace(current_space->total_array_size,¤t_space);
98: }
100: /* Copy data into free space, and zero out denserow */
101: PetscMemcpy(current_space->array,sparserow,cnzi*sizeof(int));
102: current_space->array += cnzi;
103: current_space->local_used += cnzi;
104: current_space->local_remaining -= cnzi;
105: for (j=0;j<cnzi;j++) {
106: denserow[sparserow[j]] = 0;
107: }
108: ci[i+1] = ci[i] + cnzi;
109: }
111: /* Column indices are in the list of free space */
112: /* Allocate space for cj, initialize cj, and */
113: /* destroy list of free space and other temporary array(s) */
114: PetscMalloc((ci[am]+1)*sizeof(int),&cj);
115: MakeSpaceContiguous(&free_space,cj);
116: PetscFree(denserow);
117:
118: /* Allocate space for ca */
119: PetscMalloc((ci[am]+1)*sizeof(MatScalar),&ca);
120: PetscMemzero(ca,(ci[am]+1)*sizeof(MatScalar));
121:
122: /* put together the new matrix */
123: MatCreateSeqAIJWithArrays(A->comm,am,bn,ci,cj,ca,C);
125: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
126: /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
127: c = (Mat_SeqAIJ *)((*C)->data);
128: c->freedata = PETSC_TRUE;
129: c->nonew = 0;
131: PetscLogEventEnd(logkey_matmatmult_symbolic,A,B,0,0);
132: return(0);
133: }
134: #endif /* USE_INTSORT */
136: #undef __FUNCT__
138: int MatMatMult_Symbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat *C)
139: {
140: int ierr;
141: FreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
142: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
143: int aishift=a->indexshift,bishift=b->indexshift;
144: int *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj;
145: int *ci,*cj,*lnk,idx0,idx,bcol;
146: int an=A->N,am=A->M,bn=B->N,bm=B->M;
147: int i,j,k,anzi,brow,bnzj,cnzi;
148: MatScalar *ca;
151: /* some error checking which could be moved into interface layer */
152: if (aishift || bishift) SETERRQ(PETSC_ERR_SUP,"Shifted matrix indices are not supported.");
153: if (an!=bm) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",an,bm);
154:
155: /* Set up timers */
156: if (!logkey_matmatmult_symbolic) {
157: PetscLogEventRegister(&logkey_matmatmult_symbolic,"MatMatMult_Symbolic",MAT_COOKIE);
158: }
159: PetscLogEventBegin(logkey_matmatmult_symbolic,A,B,0,0);
161: /* Set up */
162: /* Allocate ci array, arrays for fill computation and */
163: /* free space for accumulating nonzero column info */
164: PetscMalloc(((am+1)+1)*sizeof(int),&ci);
165: ci[0] = 0;
166:
167: PetscMalloc((bn+1)*sizeof(int),&lnk);
168: for (i=0; i<bn; i++) lnk[i] = -1;
170: /* Initial FreeSpace size is nnz(B)=4*bi[bm] */
171: GetMoreSpace(4*bi[bm],&free_space);
172: current_space = free_space;
174: /* Determine symbolic info for each row of the product: */
175: for (i=0;i<am;i++) {
176: anzi = ai[i+1] - ai[i];
177: cnzi = 0;
178: lnk[bn] = bn;
179: for (j=0;j<anzi;j++) {
180: brow = *aj++;
181: bnzj = bi[brow+1] - bi[brow];
182: bjj = bj + bi[brow];
183: idx = bn;
184: for (k=0;k<bnzj;k++) {
185: bcol = bjj[k];
186: if (lnk[bcol] == -1) { /* new col */
187: if (k>0) idx = bjj[k-1];
188: do {
189: idx0 = idx;
190: idx = lnk[idx0];
191: } while (bcol > idx);
192: lnk[idx0] = bcol;
193: lnk[bcol] = idx;
194: cnzi++;
195: }
196: }
197: }
199: /* If free space is not available, make more free space */
200: /* Double the amount of total space in the list */
201: if (current_space->local_remaining<cnzi) {
202: printf("...%d -th row, double space ...n",i);
203: GetMoreSpace(current_space->total_array_size,¤t_space);
204: }
206: /* Copy data into free space, and zero out denserow and lnk */
207: idx = bn;
208: for (j=0; j<cnzi; j++){
209: idx0 = idx;
210: idx = lnk[idx0];
211: *current_space->array++ = idx;
212: lnk[idx0] = -1;
213: }
214: lnk[idx] = -1;
216: current_space->local_used += cnzi;
217: current_space->local_remaining -= cnzi;
219: ci[i+1] = ci[i] + cnzi;
220: }
222: /* Column indices are in the list of free space */
223: /* Allocate space for cj, initialize cj, and */
224: /* destroy list of free space and other temporary array(s) */
225: PetscMalloc((ci[am]+1)*sizeof(int),&cj);
226: MakeSpaceContiguous(&free_space,cj);
227: PetscFree(lnk);
228:
229: /* Allocate space for ca */
230: PetscMalloc((ci[am]+1)*sizeof(MatScalar),&ca);
231: PetscMemzero(ca,(ci[am]+1)*sizeof(MatScalar));
232:
233: /* put together the new matrix */
234: MatCreateSeqAIJWithArrays(A->comm,am,bn,ci,cj,ca,C);
236: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
237: /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
238: c = (Mat_SeqAIJ *)((*C)->data);
239: c->freedata = PETSC_TRUE;
240: c->nonew = 0;
242: PetscLogEventEnd(logkey_matmatmult_symbolic,A,B,0,0);
243: return(0);
244: }
246: /*
247: MatMatMult_Numeric_SeqAIJ_SeqAIJ - Forms the numeric product of two SeqAIJ matrices
248: C=A*B;
249: Note: C must have been created by calling MatMatMult_Symbolic_SeqAIJ_SeqAIJ.
250: */
251: #undef __FUNCT__
253: int MatMatMult_Numeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C)
254: {
255: int ierr,flops=0;
256: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
257: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
258: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
259: int aishift=a->indexshift,bishift=b->indexshift,cishift=c->indexshift;
260: int *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj,*ci=c->i,*cj=c->j;
261: int an=A->N,am=A->M,bn=B->N,bm=B->M,cn=C->N,cm=C->M;
262: int i,j,k,anzi,bnzi,cnzi,brow;
263: MatScalar *aa=a->a,*ba=b->a,*baj,*ca=c->a,*temp;
267: /* This error checking should be unnecessary if the symbolic was performed */
268: if (aishift || bishift || cishift) SETERRQ(PETSC_ERR_SUP,"Shifted matrix indices are not supported.");
269: if (am!=cm) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",am,cm);
270: if (an!=bm) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",an,bm);
271: if (bn!=cn) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",bn,cn);
273: /* Set up timers */
274: if (!logkey_matmatmult_numeric) {
275: PetscLogEventRegister(&logkey_matmatmult_numeric,"MatMatMult_Numeric",MAT_COOKIE);
276: }
277: PetscLogEventBegin(logkey_matmatmult_numeric,A,B,C,0);
279: /* Allocate temp accumulation space to avoid searching for nonzero columns in C */
280: PetscMalloc((cn+1)*sizeof(MatScalar),&temp);
281: PetscMemzero(temp,cn*sizeof(MatScalar));
282: /* Traverse A row-wise. */
283: /* Build the ith row in C by summing over nonzero columns in A, */
284: /* the rows of B corresponding to nonzeros of A. */
285: for (i=0;i<am;i++) {
286: anzi = ai[i+1] - ai[i];
287: for (j=0;j<anzi;j++) {
288: brow = *aj++;
289: bnzi = bi[brow+1] - bi[brow];
290: bjj = bj + bi[brow];
291: baj = ba + bi[brow];
292: for (k=0;k<bnzi;k++) {
293: temp[bjj[k]] += (*aa)*baj[k];
294: }
295: flops += 2*bnzi;
296: aa++;
297: }
298: /* Store row back into C, and re-zero temp */
299: cnzi = ci[i+1] - ci[i];
300: for (j=0;j<cnzi;j++) {
301: ca[j] = temp[cj[j]];
302: temp[cj[j]] = 0.0;
303: }
304: ca += cnzi;
305: cj += cnzi;
306: }
307: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
308: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
309:
310: /* Free temp */
311: PetscFree(temp);
312: PetscLogFlops(flops);
313: PetscLogEventEnd(logkey_matmatmult_numeric,A,B,C,0);
314: return(0);
315: }
317: #undef __FUNCT__
319: int MatMatMult_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat *C) {
323: if (!logkey_matmatmult) {
324: PetscLogEventRegister(&logkey_matmatmult,"MatMatMult",MAT_COOKIE);
325: }
326: PetscLogEventBegin(logkey_matmatmult,A,B,0,0);
327: MatMatMult_Symbolic_SeqAIJ_SeqAIJ(A,B,C);
328: MatMatMult_Numeric_SeqAIJ_SeqAIJ(A,B,*C);
329: PetscLogEventEnd(logkey_matmatmult,A,B,0,0);
330: return(0);
331: }
334: /*
335: MatApplyPAPt_Symbolic_SeqAIJ_SeqAIJ - Forms the symbolic product of two SeqAIJ matrices
336: C = P * A * P^T;
338: Note: C is assumed to be uncreated.
339: If this is not the case, Destroy C before calling this routine.
340: */
341: #undef __FUNCT__
343: int MatApplyPAPt_Symbolic_SeqAIJ_SeqAIJ(Mat A,Mat P,Mat *C) {
344: /* Note: This code is virtually identical to that of MatApplyPtAP_SeqAIJ_Symbolic */
345: /* and MatMatMult_SeqAIJ_SeqAIJ_Symbolic. Perhaps they could be merged nicely. */
346: int ierr;
347: FreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
348: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*p=(Mat_SeqAIJ*)P->data,*c;
349: int aishift=a->indexshift,pishift=p->indexshift;
350: int *ai=a->i,*aj=a->j,*ajj,*pi=p->i,*pj=p->j,*pti,*ptj,*ptjj;
351: int *ci,*cj,*paj,*padenserow,*pasparserow,*denserow,*sparserow;
352: int an=A->N,am=A->M,pn=P->N,pm=P->M;
353: int i,j,k,pnzi,arow,anzj,panzi,ptrow,ptnzj,cnzi;
354: MatScalar *ca;
358: /* some error checking which could be moved into interface layer */
359: if (aishift || pishift) SETERRQ(PETSC_ERR_SUP,"Shifted matrix indices are not supported.");
360: if (pn!=am) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",pn,am);
361: if (am!=an) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix 'A' must be square, %d != %d",am, an);
363: /* Set up timers */
364: if (!logkey_matapplypapt_symbolic) {
365: PetscLogEventRegister(&logkey_matapplypapt_symbolic,"MatApplyPAPt_Symbolic",MAT_COOKIE);
366: }
367: PetscLogEventBegin(logkey_matapplypapt_symbolic,A,P,0,0);
369: /* Create ij structure of P^T */
370: MatGetSymbolicTranspose_SeqAIJ(P,&pti,&ptj);
372: /* Allocate ci array, arrays for fill computation and */
373: /* free space for accumulating nonzero column info */
374: PetscMalloc(((pm+1)*1)*sizeof(int),&ci);
375: ci[0] = 0;
377: PetscMalloc((2*an+2*pm+1)*sizeof(int),&padenserow);
378: PetscMemzero(padenserow,(2*an+2*pm+1)*sizeof(int));
379: pasparserow = padenserow + an;
380: denserow = pasparserow + an;
381: sparserow = denserow + pm;
383: /* Set initial free space to be nnz(A) scaled by aspect ratio of Pt. */
384: /* This should be reasonable if sparsity of PAPt is similar to that of A. */
385: ierr = GetMoreSpace((ai[am]/pn)*pm,&free_space);
386: current_space = free_space;
388: /* Determine fill for each row of C: */
389: for (i=0;i<pm;i++) {
390: pnzi = pi[i+1] - pi[i];
391: panzi = 0;
392: /* Get symbolic sparse row of PA: */
393: for (j=0;j<pnzi;j++) {
394: arow = *pj++;
395: anzj = ai[arow+1] - ai[arow];
396: ajj = aj + ai[arow];
397: for (k=0;k<anzj;k++) {
398: if (!padenserow[ajj[k]]) {
399: padenserow[ajj[k]] = -1;
400: pasparserow[panzi++] = ajj[k];
401: }
402: }
403: }
404: /* Using symbolic row of PA, determine symbolic row of C: */
405: paj = pasparserow;
406: cnzi = 0;
407: for (j=0;j<panzi;j++) {
408: ptrow = *paj++;
409: ptnzj = pti[ptrow+1] - pti[ptrow];
410: ptjj = ptj + pti[ptrow];
411: for (k=0;k<ptnzj;k++) {
412: if (!denserow[ptjj[k]]) {
413: denserow[ptjj[k]] = -1;
414: sparserow[cnzi++] = ptjj[k];
415: }
416: }
417: }
419: /* sort sparse representation */
420: PetscSortInt(cnzi,sparserow);
422: /* If free space is not available, make more free space */
423: /* Double the amount of total space in the list */
424: if (current_space->local_remaining<cnzi) {
425: GetMoreSpace(current_space->total_array_size,¤t_space);
426: }
428: /* Copy data into free space, and zero out dense row */
429: PetscMemcpy(current_space->array,sparserow,cnzi*sizeof(int));
430: current_space->array += cnzi;
431: current_space->local_used += cnzi;
432: current_space->local_remaining -= cnzi;
434: for (j=0;j<panzi;j++) {
435: padenserow[pasparserow[j]] = 0;
436: }
437: for (j=0;j<cnzi;j++) {
438: denserow[sparserow[j]] = 0;
439: }
440: ci[i+1] = ci[i] + cnzi;
441: }
442: /* column indices are in the list of free space */
443: /* Allocate space for cj, initialize cj, and */
444: /* destroy list of free space and other temporary array(s) */
445: PetscMalloc((ci[pm]+1)*sizeof(int),&cj);
446: MakeSpaceContiguous(&free_space,cj);
447: PetscFree(padenserow);
448:
449: /* Allocate space for ca */
450: PetscMalloc((ci[pm]+1)*sizeof(MatScalar),&ca);
451: PetscMemzero(ca,(ci[pm]+1)*sizeof(MatScalar));
452:
453: /* put together the new matrix */
454: MatCreateSeqAIJWithArrays(A->comm,pm,pm,ci,cj,ca,C);
456: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
457: /* Since these are PETSc arrays, change flags to free them as necessary. */
458: c = (Mat_SeqAIJ *)((*C)->data);
459: c->freedata = PETSC_TRUE;
460: c->nonew = 0;
462: /* Clean up. */
463: MatRestoreSymbolicTranspose_SeqAIJ(P,&pti,&ptj);
465: PetscLogEventEnd(logkey_matapplypapt_symbolic,A,P,0,0);
466: return(0);
467: }
469: /*
470: MatApplyPAPt_Numeric_SeqAIJ - Forms the numeric product of two SeqAIJ matrices
471: C = P * A * P^T;
472: Note: C must have been created by calling MatApplyPAPt_Symbolic_SeqAIJ.
473: */
474: #undef __FUNCT__
476: int MatApplyPAPt_Numeric_SeqAIJ_SeqAIJ(Mat A,Mat P,Mat C) {
477: int ierr,flops=0;
478: Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data;
479: Mat_SeqAIJ *p = (Mat_SeqAIJ *) P->data;
480: Mat_SeqAIJ *c = (Mat_SeqAIJ *) C->data;
481: int aishift=a->indexshift,pishift=p->indexshift,cishift=c->indexshift;
482: int *ai=a->i,*aj=a->j,*ajj,*pi=p->i,*pj=p->j,*pjj=p->j,*paj,*pajdense,*ptj;
483: int *ci=c->i,*cj=c->j;
484: int an=A->N,am=A->M,pn=P->N,pm=P->M,cn=C->N,cm=C->M;
485: int i,j,k,k1,k2,pnzi,anzj,panzj,arow,ptcol,ptnzj,cnzi;
486: MatScalar *aa=a->a,*pa=p->a,*pta=p->a,*ptaj,*paa,*aaj,*ca=c->a,sum;
490: /* This error checking should be unnecessary if the symbolic was performed */
491: if (aishift || pishift || cishift) SETERRQ(PETSC_ERR_SUP,"Shifted matrix indices are not supported.");
492: if (pm!=cm) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",pm,cm);
493: if (pn!=am) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",pn,am);
494: if (am!=an) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix 'A' must be square, %d != %d",am, an);
495: if (pm!=cn) SETERRQ2(PETSC_ERR_ARG_SIZ,"Matrix dimensions are incompatible, %d != %d",pm, cn);
497: /* Set up timers */
498: if (!logkey_matapplypapt_numeric) {
499: PetscLogEventRegister(&logkey_matapplypapt_numeric,"MatApplyPAPt_Numeric",MAT_COOKIE);
500: }
501: PetscLogEventBegin(logkey_matapplypapt_numeric,A,P,C,0);
503: PetscMalloc(an*(sizeof(MatScalar)+2*sizeof(int)),&paa);
504: PetscMemzero(paa,an*(sizeof(MatScalar)+2*sizeof(int)));
505: PetscMemzero(ca,ci[cm]*sizeof(MatScalar));
507: paj = (int *)(paa + an);
508: pajdense = paj + an;
510: for (i=0;i<pm;i++) {
511: /* Form sparse row of P*A */
512: pnzi = pi[i+1] - pi[i];
513: panzj = 0;
514: for (j=0;j<pnzi;j++) {
515: arow = *pj++;
516: anzj = ai[arow+1] - ai[arow];
517: ajj = aj + ai[arow];
518: aaj = aa + ai[arow];
519: for (k=0;k<anzj;k++) {
520: if (!pajdense[ajj[k]]) {
521: pajdense[ajj[k]] = -1;
522: paj[panzj++] = ajj[k];
523: }
524: paa[ajj[k]] += (*pa)*aaj[k];
525: }
526: flops += 2*anzj;
527: pa++;
528: }
530: /* Sort the j index array for quick sparse axpy. */
531: PetscSortInt(panzj,paj);
533: /* Compute P*A*P^T using sparse inner products. */
534: /* Take advantage of pre-computed (i,j) of C for locations of non-zeros. */
535: cnzi = ci[i+1] - ci[i];
536: for (j=0;j<cnzi;j++) {
537: /* Form sparse inner product of current row of P*A with (*cj++) col of P^T. */
538: ptcol = *cj++;
539: ptnzj = pi[ptcol+1] - pi[ptcol];
540: ptj = pjj + pi[ptcol];
541: ptaj = pta + pi[ptcol];
542: sum = 0.;
543: k1 = 0;
544: k2 = 0;
545: while ((k1<panzj) && (k2<ptnzj)) {
546: if (paj[k1]==ptj[k2]) {
547: sum += paa[paj[k1++]]*ptaj[k2++];
548: } else if (paj[k1] < ptj[k2]) {
549: k1++;
550: } else /* if (paj[k1] > ptj[k2]) */ {
551: k2++;
552: }
553: }
554: *ca++ = sum;
555: }
557: /* Zero the current row info for P*A */
558: for (j=0;j<panzj;j++) {
559: paa[paj[j]] = 0.;
560: pajdense[paj[j]] = 0;
561: }
562: }
564: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
565: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
566: PetscLogFlops(flops);
567: PetscLogEventEnd(logkey_matapplypapt_numeric,A,P,C,0);
568: return(0);
569: }
570:
571: #undef __FUNCT__
573: int MatApplyPAPt_SeqAIJ_SeqAIJ(Mat A,Mat P,Mat *C) {
577: if (!logkey_matapplypapt) {
578: PetscLogEventRegister(&logkey_matapplypapt,"MatApplyPAPt",MAT_COOKIE);
579: }
580: PetscLogEventBegin(logkey_matapplypapt,A,P,0,0);
581: MatApplyPAPt_Symbolic_SeqAIJ_SeqAIJ(A,P,C);
582: MatApplyPAPt_Numeric_SeqAIJ_SeqAIJ(A,P,*C);
583: PetscLogEventEnd(logkey_matapplypapt,A,P,0,0);
584: return(0);
585: }