Actual source code: mpidense.c
1: #define PETSCMAT_DLL
3: /*
4: Basic functions for basic parallel dense matrices.
5: */
6:
7: #include src/mat/impls/dense/mpi/mpidense.h
11: PetscErrorCode MatLUFactorSymbolic_MPIDense(Mat A,IS row,IS col,MatFactorInfo *info,Mat *fact)
12: {
16: MatDuplicate(A,MAT_DO_NOT_COPY_VALUES,fact);
17: return(0);
18: }
22: PetscErrorCode MatCholeskyFactorSymbolic_MPIDense(Mat A,IS perm,MatFactorInfo *info,Mat *fact)
23: {
27: MatDuplicate(A,MAT_DO_NOT_COPY_VALUES,fact);
28: return(0);
29: }
33: /*@
35: MatDenseGetLocalMatrix - For a MATMPIDENSE or MATSEQDENSE matrix returns the sequential
36: matrix that represents the operator. For sequential matrices it returns itself.
38: Input Parameter:
39: . A - the Seq or MPI dense matrix
41: Output Parameter:
42: . B - the inner matrix
44: Level: intermediate
46: @*/
47: PetscErrorCode MatDenseGetLocalMatrix(Mat A,Mat *B)
48: {
49: Mat_MPIDense *mat = (Mat_MPIDense*)A->data;
51: PetscTruth flg;
52: PetscMPIInt size;
55: PetscTypeCompare((PetscObject)A,MATMPIDENSE,&flg);
56: if (!flg) { /* this check sucks! */
57: PetscTypeCompare((PetscObject)A,MATDENSE,&flg);
58: if (flg) {
59: MPI_Comm_size(((PetscObject)A)->comm,&size);
60: if (size == 1) flg = PETSC_FALSE;
61: }
62: }
63: if (flg) {
64: *B = mat->A;
65: } else {
66: *B = A;
67: }
68: return(0);
69: }
73: PetscErrorCode MatGetRow_MPIDense(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
74: {
75: Mat_MPIDense *mat = (Mat_MPIDense*)A->data;
77: PetscInt lrow,rstart = A->rmap.rstart,rend = A->rmap.rend;
80: if (row < rstart || row >= rend) SETERRQ(PETSC_ERR_SUP,"only local rows")
81: lrow = row - rstart;
82: MatGetRow(mat->A,lrow,nz,(const PetscInt **)idx,(const PetscScalar **)v);
83: return(0);
84: }
88: PetscErrorCode MatRestoreRow_MPIDense(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
89: {
93: if (idx) {PetscFree(*idx);}
94: if (v) {PetscFree(*v);}
95: return(0);
96: }
101: PetscErrorCode MatGetDiagonalBlock_MPIDense(Mat A,PetscTruth *iscopy,MatReuse reuse,Mat *B)
102: {
103: Mat_MPIDense *mdn = (Mat_MPIDense*)A->data;
105: PetscInt m = A->rmap.n,rstart = A->rmap.rstart;
106: PetscScalar *array;
107: MPI_Comm comm;
110: if (A->rmap.N != A->cmap.N) SETERRQ(PETSC_ERR_SUP,"Only square matrices supported.");
112: /* The reuse aspect is not implemented efficiently */
113: if (reuse) { MatDestroy(*B);}
115: PetscObjectGetComm((PetscObject)(mdn->A),&comm);
116: MatGetArray(mdn->A,&array);
117: MatCreate(comm,B);
118: MatSetSizes(*B,m,m,m,m);
119: MatSetType(*B,((PetscObject)mdn->A)->type_name);
120: MatSeqDenseSetPreallocation(*B,array+m*rstart);
121: MatRestoreArray(mdn->A,&array);
122: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
123: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
124:
125: *iscopy = PETSC_TRUE;
126: return(0);
127: }
132: PetscErrorCode MatSetValues_MPIDense(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],const PetscScalar v[],InsertMode addv)
133: {
134: Mat_MPIDense *A = (Mat_MPIDense*)mat->data;
136: PetscInt i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend,row;
137: PetscTruth roworiented = A->roworiented;
140: for (i=0; i<m; i++) {
141: if (idxm[i] < 0) continue;
142: if (idxm[i] >= mat->rmap.N) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
143: if (idxm[i] >= rstart && idxm[i] < rend) {
144: row = idxm[i] - rstart;
145: if (roworiented) {
146: MatSetValues(A->A,1,&row,n,idxn,v+i*n,addv);
147: } else {
148: for (j=0; j<n; j++) {
149: if (idxn[j] < 0) continue;
150: if (idxn[j] >= mat->cmap.N) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
151: MatSetValues(A->A,1,&row,1,&idxn[j],v+i+j*m,addv);
152: }
153: }
154: } else {
155: if (!A->donotstash) {
156: if (roworiented) {
157: MatStashValuesRow_Private(&mat->stash,idxm[i],n,idxn,v+i*n);
158: } else {
159: MatStashValuesCol_Private(&mat->stash,idxm[i],n,idxn,v+i,m);
160: }
161: }
162: }
163: }
164: return(0);
165: }
169: PetscErrorCode MatGetValues_MPIDense(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
170: {
171: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
173: PetscInt i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend,row;
176: for (i=0; i<m; i++) {
177: if (idxm[i] < 0) continue; /* SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative row"); */
178: if (idxm[i] >= mat->rmap.N) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
179: if (idxm[i] >= rstart && idxm[i] < rend) {
180: row = idxm[i] - rstart;
181: for (j=0; j<n; j++) {
182: if (idxn[j] < 0) continue; /* SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative column"); */
183: if (idxn[j] >= mat->cmap.N) {
184: SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
185: }
186: MatGetValues(mdn->A,1,&row,1,&idxn[j],v+i*n+j);
187: }
188: } else {
189: SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
190: }
191: }
192: return(0);
193: }
197: PetscErrorCode MatGetArray_MPIDense(Mat A,PetscScalar *array[])
198: {
199: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
203: MatGetArray(a->A,array);
204: return(0);
205: }
209: static PetscErrorCode MatGetSubMatrix_MPIDense(Mat A,IS isrow,IS iscol,PetscInt cs,MatReuse scall,Mat *B)
210: {
211: Mat_MPIDense *mat = (Mat_MPIDense*)A->data,*newmatd;
212: Mat_SeqDense *lmat = (Mat_SeqDense*)mat->A->data;
214: PetscInt i,j,*irow,*icol,rstart,rend,nrows,ncols,nlrows,nlcols;
215: PetscScalar *av,*bv,*v = lmat->v;
216: Mat newmat;
219: ISGetIndices(isrow,&irow);
220: ISGetIndices(iscol,&icol);
221: ISGetLocalSize(isrow,&nrows);
222: ISGetLocalSize(iscol,&ncols);
224: /* No parallel redistribution currently supported! Should really check each index set
225: to comfirm that it is OK. ... Currently supports only submatrix same partitioning as
226: original matrix! */
228: MatGetLocalSize(A,&nlrows,&nlcols);
229: MatGetOwnershipRange(A,&rstart,&rend);
230:
231: /* Check submatrix call */
232: if (scall == MAT_REUSE_MATRIX) {
233: /* SETERRQ(PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); */
234: /* Really need to test rows and column sizes! */
235: newmat = *B;
236: } else {
237: /* Create and fill new matrix */
238: MatCreate(((PetscObject)A)->comm,&newmat);
239: MatSetSizes(newmat,nrows,cs,PETSC_DECIDE,ncols);
240: MatSetType(newmat,((PetscObject)A)->type_name);
241: MatMPIDenseSetPreallocation(newmat,PETSC_NULL);
242: }
244: /* Now extract the data pointers and do the copy, column at a time */
245: newmatd = (Mat_MPIDense*)newmat->data;
246: bv = ((Mat_SeqDense *)newmatd->A->data)->v;
247:
248: for (i=0; i<ncols; i++) {
249: av = v + nlrows*icol[i];
250: for (j=0; j<nrows; j++) {
251: *bv++ = av[irow[j] - rstart];
252: }
253: }
255: /* Assemble the matrices so that the correct flags are set */
256: MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);
257: MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);
259: /* Free work space */
260: ISRestoreIndices(isrow,&irow);
261: ISRestoreIndices(iscol,&icol);
262: *B = newmat;
263: return(0);
264: }
268: PetscErrorCode MatRestoreArray_MPIDense(Mat A,PetscScalar *array[])
269: {
271: return(0);
272: }
276: PetscErrorCode MatAssemblyBegin_MPIDense(Mat mat,MatAssemblyType mode)
277: {
278: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
279: MPI_Comm comm = ((PetscObject)mat)->comm;
281: PetscInt nstash,reallocs;
282: InsertMode addv;
285: /* make sure all processors are either in INSERTMODE or ADDMODE */
286: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,comm);
287: if (addv == (ADD_VALUES|INSERT_VALUES)) {
288: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Cannot mix adds/inserts on different procs");
289: }
290: mat->insertmode = addv; /* in case this processor had no cache */
292: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap.range);
293: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
294: PetscInfo2(mdn->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
295: return(0);
296: }
300: PetscErrorCode MatAssemblyEnd_MPIDense(Mat mat,MatAssemblyType mode)
301: {
302: Mat_MPIDense *mdn=(Mat_MPIDense*)mat->data;
303: PetscErrorCode ierr;
304: PetscInt i,*row,*col,flg,j,rstart,ncols;
305: PetscMPIInt n;
306: PetscScalar *val;
307: InsertMode addv=mat->insertmode;
310: /* wait on receives */
311: while (1) {
312: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
313: if (!flg) break;
314:
315: for (i=0; i<n;) {
316: /* Now identify the consecutive vals belonging to the same row */
317: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
318: if (j < n) ncols = j-i;
319: else ncols = n-i;
320: /* Now assemble all these values with a single function call */
321: MatSetValues_MPIDense(mat,1,row+i,ncols,col+i,val+i,addv);
322: i = j;
323: }
324: }
325: MatStashScatterEnd_Private(&mat->stash);
326:
327: MatAssemblyBegin(mdn->A,mode);
328: MatAssemblyEnd(mdn->A,mode);
330: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
331: MatSetUpMultiply_MPIDense(mat);
332: }
333: return(0);
334: }
338: PetscErrorCode MatZeroEntries_MPIDense(Mat A)
339: {
341: Mat_MPIDense *l = (Mat_MPIDense*)A->data;
344: MatZeroEntries(l->A);
345: return(0);
346: }
348: /* the code does not do the diagonal entries correctly unless the
349: matrix is square and the column and row owerships are identical.
350: This is a BUG. The only way to fix it seems to be to access
351: mdn->A and mdn->B directly and not through the MatZeroRows()
352: routine.
353: */
356: PetscErrorCode MatZeroRows_MPIDense(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag)
357: {
358: Mat_MPIDense *l = (Mat_MPIDense*)A->data;
360: PetscInt i,*owners = A->rmap.range;
361: PetscInt *nprocs,j,idx,nsends;
362: PetscInt nmax,*svalues,*starts,*owner,nrecvs;
363: PetscInt *rvalues,tag = ((PetscObject)A)->tag,count,base,slen,*source;
364: PetscInt *lens,*lrows,*values;
365: PetscMPIInt n,imdex,rank = l->rank,size = l->size;
366: MPI_Comm comm = ((PetscObject)A)->comm;
367: MPI_Request *send_waits,*recv_waits;
368: MPI_Status recv_status,*send_status;
369: PetscTruth found;
372: /* first count number of contributors to each processor */
373: PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
374: PetscMemzero(nprocs,2*size*sizeof(PetscInt));
375: PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
376: for (i=0; i<N; i++) {
377: idx = rows[i];
378: found = PETSC_FALSE;
379: for (j=0; j<size; j++) {
380: if (idx >= owners[j] && idx < owners[j+1]) {
381: nprocs[2*j]++; nprocs[2*j+1] = 1; owner[i] = j; found = PETSC_TRUE; break;
382: }
383: }
384: if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
385: }
386: nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}
388: /* inform other processors of number of messages and max length*/
389: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
391: /* post receives: */
392: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
393: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
394: for (i=0; i<nrecvs; i++) {
395: MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
396: }
398: /* do sends:
399: 1) starts[i] gives the starting index in svalues for stuff going to
400: the ith processor
401: */
402: PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
403: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
404: PetscMalloc((size+1)*sizeof(PetscInt),&starts);
405: starts[0] = 0;
406: for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
407: for (i=0; i<N; i++) {
408: svalues[starts[owner[i]]++] = rows[i];
409: }
411: starts[0] = 0;
412: for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
413: count = 0;
414: for (i=0; i<size; i++) {
415: if (nprocs[2*i+1]) {
416: MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
417: }
418: }
419: PetscFree(starts);
421: base = owners[rank];
423: /* wait on receives */
424: PetscMalloc(2*(nrecvs+1)*sizeof(PetscInt),&lens);
425: source = lens + nrecvs;
426: count = nrecvs; slen = 0;
427: while (count) {
428: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
429: /* unpack receives into our local space */
430: MPI_Get_count(&recv_status,MPIU_INT,&n);
431: source[imdex] = recv_status.MPI_SOURCE;
432: lens[imdex] = n;
433: slen += n;
434: count--;
435: }
436: PetscFree(recv_waits);
437:
438: /* move the data into the send scatter */
439: PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
440: count = 0;
441: for (i=0; i<nrecvs; i++) {
442: values = rvalues + i*nmax;
443: for (j=0; j<lens[i]; j++) {
444: lrows[count++] = values[j] - base;
445: }
446: }
447: PetscFree(rvalues);
448: PetscFree(lens);
449: PetscFree(owner);
450: PetscFree(nprocs);
451:
452: /* actually zap the local rows */
453: MatZeroRows(l->A,slen,lrows,diag);
454: PetscFree(lrows);
456: /* wait on sends */
457: if (nsends) {
458: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
459: MPI_Waitall(nsends,send_waits,send_status);
460: PetscFree(send_status);
461: }
462: PetscFree(send_waits);
463: PetscFree(svalues);
465: return(0);
466: }
470: PetscErrorCode MatMult_MPIDense(Mat mat,Vec xx,Vec yy)
471: {
472: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
476: VecScatterBegin(mdn->Mvctx,xx,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
477: VecScatterEnd(mdn->Mvctx,xx,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
478: MatMult_SeqDense(mdn->A,mdn->lvec,yy);
479: return(0);
480: }
484: PetscErrorCode MatMultAdd_MPIDense(Mat mat,Vec xx,Vec yy,Vec zz)
485: {
486: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
490: VecScatterBegin(mdn->Mvctx,xx,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
491: VecScatterEnd(mdn->Mvctx,xx,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
492: MatMultAdd_SeqDense(mdn->A,mdn->lvec,yy,zz);
493: return(0);
494: }
498: PetscErrorCode MatMultTranspose_MPIDense(Mat A,Vec xx,Vec yy)
499: {
500: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
502: PetscScalar zero = 0.0;
505: VecSet(yy,zero);
506: MatMultTranspose_SeqDense(a->A,xx,a->lvec);
507: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
508: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
509: return(0);
510: }
514: PetscErrorCode MatMultTransposeAdd_MPIDense(Mat A,Vec xx,Vec yy,Vec zz)
515: {
516: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
520: VecCopy(yy,zz);
521: MatMultTranspose_SeqDense(a->A,xx,a->lvec);
522: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
523: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
524: return(0);
525: }
529: PetscErrorCode MatGetDiagonal_MPIDense(Mat A,Vec v)
530: {
531: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
532: Mat_SeqDense *aloc = (Mat_SeqDense*)a->A->data;
534: PetscInt len,i,n,m = A->rmap.n,radd;
535: PetscScalar *x,zero = 0.0;
536:
538: VecSet(v,zero);
539: VecGetArray(v,&x);
540: VecGetSize(v,&n);
541: if (n != A->rmap.N) SETERRQ(PETSC_ERR_ARG_SIZ,"Nonconforming mat and vec");
542: len = PetscMin(a->A->rmap.n,a->A->cmap.n);
543: radd = A->rmap.rstart*m;
544: for (i=0; i<len; i++) {
545: x[i] = aloc->v[radd + i*m + i];
546: }
547: VecRestoreArray(v,&x);
548: return(0);
549: }
553: PetscErrorCode MatDestroy_MPIDense(Mat mat)
554: {
555: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
560: #if defined(PETSC_USE_LOG)
561: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap.N,mat->cmap.N);
562: #endif
563: MatStashDestroy_Private(&mat->stash);
564: MatDestroy(mdn->A);
565: if (mdn->lvec) {VecDestroy(mdn->lvec);}
566: if (mdn->Mvctx) {VecScatterDestroy(mdn->Mvctx);}
567: if (mdn->factor) {
568: PetscFree(mdn->factor->temp);
569: PetscFree(mdn->factor->tag);
570: PetscFree(mdn->factor->pivots);
571: PetscFree(mdn->factor);
572: }
573: PetscFree(mdn);
574: PetscObjectChangeTypeName((PetscObject)mat,0);
575: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
576: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatMPIDenseSetPreallocation_C","",PETSC_NULL);
577: PetscObjectComposeFunction((PetscObject)mat,"MatMatMult_mpiaij_mpidense_C","",PETSC_NULL);
578: PetscObjectComposeFunction((PetscObject)mat,"MatMatMultSymbolic_mpiaij_mpidense_C","",PETSC_NULL);
579: PetscObjectComposeFunction((PetscObject)mat,"MatMatMultNumeric_mpiaij_mpidense_C","",PETSC_NULL);
580: return(0);
581: }
585: static PetscErrorCode MatView_MPIDense_Binary(Mat mat,PetscViewer viewer)
586: {
587: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
588: PetscErrorCode ierr;
589: PetscViewerFormat format;
590: int fd;
591: PetscInt header[4],mmax,N = mat->cmap.N,i,j,m,k;
592: PetscMPIInt rank,tag = ((PetscObject)viewer)->tag,size;
593: PetscScalar *work,*v,*vv;
594: Mat_SeqDense *a = (Mat_SeqDense*)mdn->A->data;
595: MPI_Status status;
598: if (mdn->size == 1) {
599: MatView(mdn->A,viewer);
600: } else {
601: PetscViewerBinaryGetDescriptor(viewer,&fd);
602: MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
603: MPI_Comm_size(((PetscObject)mat)->comm,&size);
605: PetscViewerGetFormat(viewer,&format);
606: if (format == PETSC_VIEWER_BINARY_NATIVE) {
608: if (!rank) {
609: /* store the matrix as a dense matrix */
610: header[0] = MAT_FILE_COOKIE;
611: header[1] = mat->rmap.N;
612: header[2] = N;
613: header[3] = MATRIX_BINARY_FORMAT_DENSE;
614: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
616: /* get largest work array needed for transposing array */
617: mmax = mat->rmap.n;
618: for (i=1; i<size; i++) {
619: mmax = PetscMax(mmax,mat->rmap.range[i+1] - mat->rmap.range[i]);
620: }
621: PetscMalloc(mmax*N*sizeof(PetscScalar),&work);
623: /* write out local array, by rows */
624: m = mat->rmap.n;
625: v = a->v;
626: for (j=0; j<N; j++) {
627: for (i=0; i<m; i++) {
628: work[j + i*N] = *v++;
629: }
630: }
631: PetscBinaryWrite(fd,work,m*N,PETSC_SCALAR,PETSC_FALSE);
632: /* get largest work array to receive messages from other processes, excludes process zero */
633: mmax = 0;
634: for (i=1; i<size; i++) {
635: mmax = PetscMax(mmax,mat->rmap.range[i+1] - mat->rmap.range[i]);
636: }
637: PetscMalloc(mmax*N*sizeof(PetscScalar),&vv);
638: v = vv;
639: for (k=1; k<size; k++) {
640: m = mat->rmap.range[k+1] - mat->rmap.range[k];
641: MPI_Recv(v,m*N,MPIU_SCALAR,k,tag,((PetscObject)mat)->comm,&status);
643: for (j=0; j<N; j++) {
644: for (i=0; i<m; i++) {
645: work[j + i*N] = *v++;
646: }
647: }
648: PetscBinaryWrite(fd,work,m*N,PETSC_SCALAR,PETSC_FALSE);
649: }
650: PetscFree(work);
651: PetscFree(vv);
652: } else {
653: MPI_Send(a->v,mat->rmap.n*mat->cmap.N,MPIU_SCALAR,0,tag,((PetscObject)mat)->comm);
654: }
655: }
656: }
657: return(0);
658: }
662: static PetscErrorCode MatView_MPIDense_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
663: {
664: Mat_MPIDense *mdn = (Mat_MPIDense*)mat->data;
665: PetscErrorCode ierr;
666: PetscMPIInt size = mdn->size,rank = mdn->rank;
667: PetscViewerType vtype;
668: PetscTruth iascii,isdraw;
669: PetscViewer sviewer;
670: PetscViewerFormat format;
673: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
674: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
675: if (iascii) {
676: PetscViewerGetType(viewer,&vtype);
677: PetscViewerGetFormat(viewer,&format);
678: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
679: MatInfo info;
680: MatGetInfo(mat,MAT_LOCAL,&info);
681: PetscViewerASCIISynchronizedPrintf(viewer," [%d] local rows %D nz %D nz alloced %D mem %D \n",rank,mat->rmap.n,
682: (PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
683: PetscViewerFlush(viewer);
684: VecScatterView(mdn->Mvctx,viewer);
685: return(0);
686: } else if (format == PETSC_VIEWER_ASCII_INFO) {
687: return(0);
688: }
689: } else if (isdraw) {
690: PetscDraw draw;
691: PetscTruth isnull;
693: PetscViewerDrawGetDraw(viewer,0,&draw);
694: PetscDrawIsNull(draw,&isnull);
695: if (isnull) return(0);
696: }
698: if (size == 1) {
699: MatView(mdn->A,viewer);
700: } else {
701: /* assemble the entire matrix onto first processor. */
702: Mat A;
703: PetscInt M = mat->rmap.N,N = mat->cmap.N,m,row,i,nz;
704: PetscInt *cols;
705: PetscScalar *vals;
707: MatCreate(((PetscObject)mat)->comm,&A);
708: if (!rank) {
709: MatSetSizes(A,M,N,M,N);
710: } else {
711: MatSetSizes(A,0,0,M,N);
712: }
713: /* Since this is a temporary matrix, MATMPIDENSE instead of ((PetscObject)A)->type_name here is probably acceptable. */
714: MatSetType(A,MATMPIDENSE);
715: MatMPIDenseSetPreallocation(A,PETSC_NULL);
716: PetscLogObjectParent(mat,A);
718: /* Copy the matrix ... This isn't the most efficient means,
719: but it's quick for now */
720: A->insertmode = INSERT_VALUES;
721: row = mat->rmap.rstart; m = mdn->A->rmap.n;
722: for (i=0; i<m; i++) {
723: MatGetRow_MPIDense(mat,row,&nz,&cols,&vals);
724: MatSetValues_MPIDense(A,1,&row,nz,cols,vals,INSERT_VALUES);
725: MatRestoreRow_MPIDense(mat,row,&nz,&cols,&vals);
726: row++;
727: }
729: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
730: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
731: PetscViewerGetSingleton(viewer,&sviewer);
732: if (!rank) {
733: MatView(((Mat_MPIDense*)(A->data))->A,sviewer);
734: }
735: PetscViewerRestoreSingleton(viewer,&sviewer);
736: PetscViewerFlush(viewer);
737: MatDestroy(A);
738: }
739: return(0);
740: }
744: PetscErrorCode MatView_MPIDense(Mat mat,PetscViewer viewer)
745: {
747: PetscTruth iascii,isbinary,isdraw,issocket;
748:
750:
751: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
752: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
753: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
754: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
756: if (iascii || issocket || isdraw) {
757: MatView_MPIDense_ASCIIorDraworSocket(mat,viewer);
758: } else if (isbinary) {
759: MatView_MPIDense_Binary(mat,viewer);
760: } else {
761: SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPI dense matrix",((PetscObject)viewer)->type_name);
762: }
763: return(0);
764: }
768: PetscErrorCode MatGetInfo_MPIDense(Mat A,MatInfoType flag,MatInfo *info)
769: {
770: Mat_MPIDense *mat = (Mat_MPIDense*)A->data;
771: Mat mdn = mat->A;
773: PetscReal isend[5],irecv[5];
776: info->rows_global = (double)A->rmap.N;
777: info->columns_global = (double)A->cmap.N;
778: info->rows_local = (double)A->rmap.n;
779: info->columns_local = (double)A->cmap.N;
780: info->block_size = 1.0;
781: MatGetInfo(mdn,MAT_LOCAL,info);
782: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
783: isend[3] = info->memory; isend[4] = info->mallocs;
784: if (flag == MAT_LOCAL) {
785: info->nz_used = isend[0];
786: info->nz_allocated = isend[1];
787: info->nz_unneeded = isend[2];
788: info->memory = isend[3];
789: info->mallocs = isend[4];
790: } else if (flag == MAT_GLOBAL_MAX) {
791: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,((PetscObject)A)->comm);
792: info->nz_used = irecv[0];
793: info->nz_allocated = irecv[1];
794: info->nz_unneeded = irecv[2];
795: info->memory = irecv[3];
796: info->mallocs = irecv[4];
797: } else if (flag == MAT_GLOBAL_SUM) {
798: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,((PetscObject)A)->comm);
799: info->nz_used = irecv[0];
800: info->nz_allocated = irecv[1];
801: info->nz_unneeded = irecv[2];
802: info->memory = irecv[3];
803: info->mallocs = irecv[4];
804: }
805: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
806: info->fill_ratio_needed = 0;
807: info->factor_mallocs = 0;
808: return(0);
809: }
813: PetscErrorCode MatSetOption_MPIDense(Mat A,MatOption op,PetscTruth flg)
814: {
815: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
819: switch (op) {
820: case MAT_NEW_NONZERO_LOCATIONS:
821: case MAT_NEW_NONZERO_LOCATION_ERR:
822: case MAT_NEW_NONZERO_ALLOCATION_ERR:
823: MatSetOption(a->A,op,flg);
824: break;
825: case MAT_ROW_ORIENTED:
826: a->roworiented = flg;
827: MatSetOption(a->A,op,flg);
828: break;
829: case MAT_NEW_DIAGONALS:
830: case MAT_USE_HASH_TABLE:
831: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
832: break;
833: case MAT_IGNORE_OFF_PROC_ENTRIES:
834: a->donotstash = flg;
835: break;
836: case MAT_SYMMETRIC:
837: case MAT_STRUCTURALLY_SYMMETRIC:
838: case MAT_HERMITIAN:
839: case MAT_SYMMETRY_ETERNAL:
840: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
841: break;
842: default:
843: SETERRQ1(PETSC_ERR_SUP,"unknown option %d",op);
844: }
845: return(0);
846: }
851: PetscErrorCode MatDiagonalScale_MPIDense(Mat A,Vec ll,Vec rr)
852: {
853: Mat_MPIDense *mdn = (Mat_MPIDense*)A->data;
854: Mat_SeqDense *mat = (Mat_SeqDense*)mdn->A->data;
855: PetscScalar *l,*r,x,*v;
857: PetscInt i,j,s2a,s3a,s2,s3,m=mdn->A->rmap.n,n=mdn->A->cmap.n;
860: MatGetLocalSize(A,&s2,&s3);
861: if (ll) {
862: VecGetLocalSize(ll,&s2a);
863: if (s2a != s2) SETERRQ2(PETSC_ERR_ARG_SIZ,"Left scaling vector non-conforming local size, %d != %d.", s2a, s2);
864: VecGetArray(ll,&l);
865: for (i=0; i<m; i++) {
866: x = l[i];
867: v = mat->v + i;
868: for (j=0; j<n; j++) { (*v) *= x; v+= m;}
869: }
870: VecRestoreArray(ll,&l);
871: PetscLogFlops(n*m);
872: }
873: if (rr) {
874: VecGetLocalSize(rr,&s3a);
875: if (s3a != s3) SETERRQ2(PETSC_ERR_ARG_SIZ,"Right scaling vec non-conforming local size, %d != %d.", s3a, s3);
876: VecScatterBegin(mdn->Mvctx,rr,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
877: VecScatterEnd(mdn->Mvctx,rr,mdn->lvec,INSERT_VALUES,SCATTER_FORWARD);
878: VecGetArray(mdn->lvec,&r);
879: for (i=0; i<n; i++) {
880: x = r[i];
881: v = mat->v + i*m;
882: for (j=0; j<m; j++) { (*v++) *= x;}
883: }
884: VecRestoreArray(mdn->lvec,&r);
885: PetscLogFlops(n*m);
886: }
887: return(0);
888: }
892: PetscErrorCode MatNorm_MPIDense(Mat A,NormType type,PetscReal *nrm)
893: {
894: Mat_MPIDense *mdn = (Mat_MPIDense*)A->data;
895: Mat_SeqDense *mat = (Mat_SeqDense*)mdn->A->data;
897: PetscInt i,j;
898: PetscReal sum = 0.0;
899: PetscScalar *v = mat->v;
902: if (mdn->size == 1) {
903: MatNorm(mdn->A,type,nrm);
904: } else {
905: if (type == NORM_FROBENIUS) {
906: for (i=0; i<mdn->A->cmap.n*mdn->A->rmap.n; i++) {
907: #if defined(PETSC_USE_COMPLEX)
908: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
909: #else
910: sum += (*v)*(*v); v++;
911: #endif
912: }
913: MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPI_SUM,((PetscObject)A)->comm);
914: *nrm = sqrt(*nrm);
915: PetscLogFlops(2*mdn->A->cmap.n*mdn->A->rmap.n);
916: } else if (type == NORM_1) {
917: PetscReal *tmp,*tmp2;
918: PetscMalloc(2*A->cmap.N*sizeof(PetscReal),&tmp);
919: tmp2 = tmp + A->cmap.N;
920: PetscMemzero(tmp,2*A->cmap.N*sizeof(PetscReal));
921: *nrm = 0.0;
922: v = mat->v;
923: for (j=0; j<mdn->A->cmap.n; j++) {
924: for (i=0; i<mdn->A->rmap.n; i++) {
925: tmp[j] += PetscAbsScalar(*v); v++;
926: }
927: }
928: MPI_Allreduce(tmp,tmp2,A->cmap.N,MPIU_REAL,MPI_SUM,((PetscObject)A)->comm);
929: for (j=0; j<A->cmap.N; j++) {
930: if (tmp2[j] > *nrm) *nrm = tmp2[j];
931: }
932: PetscFree(tmp);
933: PetscLogFlops(A->cmap.n*A->rmap.n);
934: } else if (type == NORM_INFINITY) { /* max row norm */
935: PetscReal ntemp;
936: MatNorm(mdn->A,type,&ntemp);
937: MPI_Allreduce(&ntemp,nrm,1,MPIU_REAL,MPI_MAX,((PetscObject)A)->comm);
938: } else {
939: SETERRQ(PETSC_ERR_SUP,"No support for two norm");
940: }
941: }
942: return(0);
943: }
947: PetscErrorCode MatTranspose_MPIDense(Mat A,Mat *matout)
948: {
949: Mat_MPIDense *a = (Mat_MPIDense*)A->data;
950: Mat_SeqDense *Aloc = (Mat_SeqDense*)a->A->data;
951: Mat B;
952: PetscInt M = A->rmap.N,N = A->cmap.N,m,n,*rwork,rstart = A->rmap.rstart;
954: PetscInt j,i;
955: PetscScalar *v;
958: if (!matout && M != N) {
959: SETERRQ(PETSC_ERR_SUP,"Supports square matrix only in-place");
960: }
961: MatCreate(((PetscObject)A)->comm,&B);
962: MatSetSizes(B,PETSC_DECIDE,PETSC_DECIDE,N,M);
963: MatSetType(B,((PetscObject)A)->type_name);
964: MatMPIDenseSetPreallocation(B,PETSC_NULL);
966: m = a->A->rmap.n; n = a->A->cmap.n; v = Aloc->v;
967: PetscMalloc(m*sizeof(PetscInt),&rwork);
968: for (i=0; i<m; i++) rwork[i] = rstart + i;
969: for (j=0; j<n; j++) {
970: MatSetValues(B,1,&j,m,rwork,v,INSERT_VALUES);
971: v += m;
972: }
973: PetscFree(rwork);
974: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
975: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
976: if (matout) {
977: *matout = B;
978: } else {
979: MatHeaderCopy(A,B);
980: }
981: return(0);
982: }
984: #include petscblaslapack.h
987: PetscErrorCode MatScale_MPIDense(Mat inA,PetscScalar alpha)
988: {
989: Mat_MPIDense *A = (Mat_MPIDense*)inA->data;
990: Mat_SeqDense *a = (Mat_SeqDense*)A->A->data;
991: PetscScalar oalpha = alpha;
992: PetscBLASInt one = 1,nz = (PetscBLASInt)inA->rmap.n*inA->cmap.N;
996: BLASscal_(&nz,&oalpha,a->v,&one);
997: PetscLogFlops(nz);
998: return(0);
999: }
1001: static PetscErrorCode MatDuplicate_MPIDense(Mat,MatDuplicateOption,Mat *);
1005: PetscErrorCode MatSetUpPreallocation_MPIDense(Mat A)
1006: {
1010: MatMPIDenseSetPreallocation(A,0);
1011: return(0);
1012: }
1016: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C)
1017: {
1019: PetscInt m=A->rmap.n,n=B->cmap.n;
1020: Mat Cmat;
1023: if (A->cmap.n != B->rmap.n) SETERRQ2(PETSC_ERR_ARG_SIZ,"A->cmap.n %d != B->rmap.n %d\n",A->cmap.n,B->rmap.n);
1024: MatCreate(((PetscObject)B)->comm,&Cmat);
1025: MatSetSizes(Cmat,m,n,A->rmap.N,B->cmap.N);
1026: MatSetType(Cmat,MATMPIDENSE);
1027: MatMPIDenseSetPreallocation(Cmat,PETSC_NULL);
1028: MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
1029: MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
1030: *C = Cmat;
1031: return(0);
1032: }
1034: /* -------------------------------------------------------------------*/
1035: static struct _MatOps MatOps_Values = {MatSetValues_MPIDense,
1036: MatGetRow_MPIDense,
1037: MatRestoreRow_MPIDense,
1038: MatMult_MPIDense,
1039: /* 4*/ MatMultAdd_MPIDense,
1040: MatMultTranspose_MPIDense,
1041: MatMultTransposeAdd_MPIDense,
1042: 0,
1043: 0,
1044: 0,
1045: /*10*/ 0,
1046: 0,
1047: 0,
1048: 0,
1049: MatTranspose_MPIDense,
1050: /*15*/ MatGetInfo_MPIDense,
1051: MatEqual_MPIDense,
1052: MatGetDiagonal_MPIDense,
1053: MatDiagonalScale_MPIDense,
1054: MatNorm_MPIDense,
1055: /*20*/ MatAssemblyBegin_MPIDense,
1056: MatAssemblyEnd_MPIDense,
1057: 0,
1058: MatSetOption_MPIDense,
1059: MatZeroEntries_MPIDense,
1060: /*25*/ MatZeroRows_MPIDense,
1061: MatLUFactorSymbolic_MPIDense,
1062: 0,
1063: MatCholeskyFactorSymbolic_MPIDense,
1064: 0,
1065: /*30*/ MatSetUpPreallocation_MPIDense,
1066: 0,
1067: 0,
1068: MatGetArray_MPIDense,
1069: MatRestoreArray_MPIDense,
1070: /*35*/ MatDuplicate_MPIDense,
1071: 0,
1072: 0,
1073: 0,
1074: 0,
1075: /*40*/ 0,
1076: MatGetSubMatrices_MPIDense,
1077: 0,
1078: MatGetValues_MPIDense,
1079: 0,
1080: /*45*/ 0,
1081: MatScale_MPIDense,
1082: 0,
1083: 0,
1084: 0,
1085: /*50*/ 0,
1086: 0,
1087: 0,
1088: 0,
1089: 0,
1090: /*55*/ 0,
1091: 0,
1092: 0,
1093: 0,
1094: 0,
1095: /*60*/ MatGetSubMatrix_MPIDense,
1096: MatDestroy_MPIDense,
1097: MatView_MPIDense,
1098: 0,
1099: 0,
1100: /*65*/ 0,
1101: 0,
1102: 0,
1103: 0,
1104: 0,
1105: /*70*/ 0,
1106: 0,
1107: 0,
1108: 0,
1109: 0,
1110: /*75*/ 0,
1111: 0,
1112: 0,
1113: 0,
1114: 0,
1115: /*80*/ 0,
1116: 0,
1117: 0,
1118: 0,
1119: /*84*/ MatLoad_MPIDense,
1120: 0,
1121: 0,
1122: 0,
1123: 0,
1124: 0,
1125: /*90*/ 0,
1126: MatMatMultSymbolic_MPIDense_MPIDense,
1127: 0,
1128: 0,
1129: 0,
1130: /*95*/ 0,
1131: 0,
1132: 0,
1133: 0};
1138: PetscErrorCode MatMPIDenseSetPreallocation_MPIDense(Mat mat,PetscScalar *data)
1139: {
1140: Mat_MPIDense *a;
1144: mat->preallocated = PETSC_TRUE;
1145: /* Note: For now, when data is specified above, this assumes the user correctly
1146: allocates the local dense storage space. We should add error checking. */
1148: a = (Mat_MPIDense*)mat->data;
1149: MatCreate(PETSC_COMM_SELF,&a->A);
1150: MatSetSizes(a->A,mat->rmap.n,mat->cmap.N,mat->rmap.n,mat->cmap.N);
1151: MatSetType(a->A,MATSEQDENSE);
1152: MatSeqDenseSetPreallocation(a->A,data);
1153: PetscLogObjectParent(mat,a->A);
1154: return(0);
1155: }
1158: /*MC
1159: MATMPIDENSE - MATMPIDENSE = "mpidense" - A matrix type to be used for distributed dense matrices.
1161: Options Database Keys:
1162: . -mat_type mpidense - sets the matrix type to "mpidense" during a call to MatSetFromOptions()
1164: Level: beginner
1166: .seealso: MatCreateMPIDense
1167: M*/
1172: PetscErrorCode MatCreate_MPIDense(Mat mat)
1173: {
1174: Mat_MPIDense *a;
1178: PetscNewLog(mat,Mat_MPIDense,&a);
1179: mat->data = (void*)a;
1180: PetscMemcpy(mat->ops,&MatOps_Values,sizeof(struct _MatOps));
1181: mat->factor = 0;
1182: mat->mapping = 0;
1184: a->factor = 0;
1185: mat->insertmode = NOT_SET_VALUES;
1186: MPI_Comm_rank(((PetscObject)mat)->comm,&a->rank);
1187: MPI_Comm_size(((PetscObject)mat)->comm,&a->size);
1189: mat->rmap.bs = mat->cmap.bs = 1;
1190: PetscMapSetUp(&mat->rmap);
1191: PetscMapSetUp(&mat->cmap);
1192: a->nvec = mat->cmap.n;
1194: /* build cache for off array entries formed */
1195: a->donotstash = PETSC_FALSE;
1196: MatStashCreate_Private(((PetscObject)mat)->comm,1,&mat->stash);
1198: /* stuff used for matrix vector multiply */
1199: a->lvec = 0;
1200: a->Mvctx = 0;
1201: a->roworiented = PETSC_TRUE;
1203: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatGetDiagonalBlock_C",
1204: "MatGetDiagonalBlock_MPIDense",
1205: MatGetDiagonalBlock_MPIDense);
1206: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatMPIDenseSetPreallocation_C",
1207: "MatMPIDenseSetPreallocation_MPIDense",
1208: MatMPIDenseSetPreallocation_MPIDense);
1209: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatMatMult_mpiaij_mpidense_C",
1210: "MatMatMult_MPIAIJ_MPIDense",
1211: MatMatMult_MPIAIJ_MPIDense);
1212: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatMatMultSymbolic_mpiaij_mpidense_C",
1213: "MatMatMultSymbolic_MPIAIJ_MPIDense",
1214: MatMatMultSymbolic_MPIAIJ_MPIDense);
1215: PetscObjectComposeFunctionDynamic((PetscObject)mat,"MatMatMultNumeric_mpiaij_mpidense_C",
1216: "MatMatMultNumeric_MPIAIJ_MPIDense",
1217: MatMatMultNumeric_MPIAIJ_MPIDense);
1218: PetscObjectChangeTypeName((PetscObject)mat,MATMPIDENSE);
1219: return(0);
1220: }
1223: /*MC
1224: MATDENSE - MATDENSE = "dense" - A matrix type to be used for dense matrices.
1226: This matrix type is identical to MATSEQDENSE when constructed with a single process communicator,
1227: and MATMPIDENSE otherwise.
1229: Options Database Keys:
1230: . -mat_type dense - sets the matrix type to "dense" during a call to MatSetFromOptions()
1232: Level: beginner
1234: .seealso: MatCreateMPIDense,MATSEQDENSE,MATMPIDENSE
1235: M*/
1240: PetscErrorCode MatCreate_Dense(Mat A)
1241: {
1243: PetscMPIInt size;
1246: MPI_Comm_size(((PetscObject)A)->comm,&size);
1247: if (size == 1) {
1248: MatSetType(A,MATSEQDENSE);
1249: } else {
1250: MatSetType(A,MATMPIDENSE);
1251: }
1252: return(0);
1253: }
1258: /*@C
1259: MatMPIDenseSetPreallocation - Sets the array used to store the matrix entries
1261: Not collective
1263: Input Parameters:
1264: . A - the matrix
1265: - data - optional location of matrix data. Set data=PETSC_NULL for PETSc
1266: to control all matrix memory allocation.
1268: Notes:
1269: The dense format is fully compatible with standard Fortran 77
1270: storage by columns.
1272: The data input variable is intended primarily for Fortran programmers
1273: who wish to allocate their own matrix memory space. Most users should
1274: set data=PETSC_NULL.
1276: Level: intermediate
1278: .keywords: matrix,dense, parallel
1280: .seealso: MatCreate(), MatCreateSeqDense(), MatSetValues()
1281: @*/
1282: PetscErrorCode MatMPIDenseSetPreallocation(Mat mat,PetscScalar *data)
1283: {
1284: PetscErrorCode ierr,(*f)(Mat,PetscScalar *);
1287: PetscObjectQueryFunction((PetscObject)mat,"MatMPIDenseSetPreallocation_C",(void (**)(void))&f);
1288: if (f) {
1289: (*f)(mat,data);
1290: }
1291: return(0);
1292: }
1296: /*@C
1297: MatCreateMPIDense - Creates a sparse parallel matrix in dense format.
1299: Collective on MPI_Comm
1301: Input Parameters:
1302: + comm - MPI communicator
1303: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
1304: . n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
1305: . M - number of global rows (or PETSC_DECIDE to have calculated if m is given)
1306: . N - number of global columns (or PETSC_DECIDE to have calculated if n is given)
1307: - data - optional location of matrix data. Set data=PETSC_NULL (PETSC_NULL_SCALAR for Fortran users) for PETSc
1308: to control all matrix memory allocation.
1310: Output Parameter:
1311: . A - the matrix
1313: Notes:
1314: The dense format is fully compatible with standard Fortran 77
1315: storage by columns.
1317: The data input variable is intended primarily for Fortran programmers
1318: who wish to allocate their own matrix memory space. Most users should
1319: set data=PETSC_NULL (PETSC_NULL_SCALAR for Fortran users).
1321: The user MUST specify either the local or global matrix dimensions
1322: (possibly both).
1324: Level: intermediate
1326: .keywords: matrix,dense, parallel
1328: .seealso: MatCreate(), MatCreateSeqDense(), MatSetValues()
1329: @*/
1330: PetscErrorCode MatCreateMPIDense(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscScalar *data,Mat *A)
1331: {
1333: PetscMPIInt size;
1336: MatCreate(comm,A);
1337: MatSetSizes(*A,m,n,M,N);
1338: MPI_Comm_size(comm,&size);
1339: if (size > 1) {
1340: MatSetType(*A,MATMPIDENSE);
1341: MatMPIDenseSetPreallocation(*A,data);
1342: } else {
1343: MatSetType(*A,MATSEQDENSE);
1344: MatSeqDenseSetPreallocation(*A,data);
1345: }
1346: return(0);
1347: }
1351: static PetscErrorCode MatDuplicate_MPIDense(Mat A,MatDuplicateOption cpvalues,Mat *newmat)
1352: {
1353: Mat mat;
1354: Mat_MPIDense *a,*oldmat = (Mat_MPIDense*)A->data;
1358: *newmat = 0;
1359: MatCreate(((PetscObject)A)->comm,&mat);
1360: MatSetSizes(mat,A->rmap.n,A->cmap.n,A->rmap.N,A->cmap.N);
1361: MatSetType(mat,((PetscObject)A)->type_name);
1362: a = (Mat_MPIDense*)mat->data;
1363: PetscMemcpy(mat->ops,A->ops,sizeof(struct _MatOps));
1364: mat->factor = A->factor;
1365: mat->assembled = PETSC_TRUE;
1366: mat->preallocated = PETSC_TRUE;
1368: mat->rmap.rstart = A->rmap.rstart;
1369: mat->rmap.rend = A->rmap.rend;
1370: a->size = oldmat->size;
1371: a->rank = oldmat->rank;
1372: mat->insertmode = NOT_SET_VALUES;
1373: a->nvec = oldmat->nvec;
1374: a->donotstash = oldmat->donotstash;
1375:
1376: PetscMemcpy(mat->rmap.range,A->rmap.range,(a->size+1)*sizeof(PetscInt));
1377: PetscMemcpy(mat->cmap.range,A->cmap.range,(a->size+1)*sizeof(PetscInt));
1378: MatStashCreate_Private(((PetscObject)A)->comm,1,&mat->stash);
1380: MatSetUpMultiply_MPIDense(mat);
1381: MatDuplicate(oldmat->A,cpvalues,&a->A);
1382: PetscLogObjectParent(mat,a->A);
1383: *newmat = mat;
1384: return(0);
1385: }
1387: #include petscsys.h
1391: PetscErrorCode MatLoad_MPIDense_DenseInFile(MPI_Comm comm,PetscInt fd,PetscInt M,PetscInt N, MatType type,Mat *newmat)
1392: {
1394: PetscMPIInt rank,size;
1395: PetscInt *rowners,i,m,nz,j;
1396: PetscScalar *array,*vals,*vals_ptr;
1397: MPI_Status status;
1400: MPI_Comm_rank(comm,&rank);
1401: MPI_Comm_size(comm,&size);
1403: /* determine ownership of all rows */
1404: m = M/size + ((M % size) > rank);
1405: PetscMalloc((size+2)*sizeof(PetscInt),&rowners);
1406: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
1407: rowners[0] = 0;
1408: for (i=2; i<=size; i++) {
1409: rowners[i] += rowners[i-1];
1410: }
1412: MatCreate(comm,newmat);
1413: MatSetSizes(*newmat,m,PETSC_DECIDE,M,N);
1414: MatSetType(*newmat,type);
1415: MatMPIDenseSetPreallocation(*newmat,PETSC_NULL);
1416: MatGetArray(*newmat,&array);
1418: if (!rank) {
1419: PetscMalloc(m*N*sizeof(PetscScalar),&vals);
1421: /* read in my part of the matrix numerical values */
1422: PetscBinaryRead(fd,vals,m*N,PETSC_SCALAR);
1423:
1424: /* insert into matrix-by row (this is why cannot directly read into array */
1425: vals_ptr = vals;
1426: for (i=0; i<m; i++) {
1427: for (j=0; j<N; j++) {
1428: array[i + j*m] = *vals_ptr++;
1429: }
1430: }
1432: /* read in other processors and ship out */
1433: for (i=1; i<size; i++) {
1434: nz = (rowners[i+1] - rowners[i])*N;
1435: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1436: MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)(*newmat))->tag,comm);
1437: }
1438: } else {
1439: /* receive numeric values */
1440: PetscMalloc(m*N*sizeof(PetscScalar),&vals);
1442: /* receive message of values*/
1443: MPI_Recv(vals,m*N,MPIU_SCALAR,0,((PetscObject)(*newmat))->tag,comm,&status);
1445: /* insert into matrix-by row (this is why cannot directly read into array */
1446: vals_ptr = vals;
1447: for (i=0; i<m; i++) {
1448: for (j=0; j<N; j++) {
1449: array[i + j*m] = *vals_ptr++;
1450: }
1451: }
1452: }
1453: PetscFree(rowners);
1454: PetscFree(vals);
1455: MatAssemblyBegin(*newmat,MAT_FINAL_ASSEMBLY);
1456: MatAssemblyEnd(*newmat,MAT_FINAL_ASSEMBLY);
1457: return(0);
1458: }
1462: PetscErrorCode MatLoad_MPIDense(PetscViewer viewer, MatType type,Mat *newmat)
1463: {
1464: Mat A;
1465: PetscScalar *vals,*svals;
1466: MPI_Comm comm = ((PetscObject)viewer)->comm;
1467: MPI_Status status;
1468: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag,*rowners,*sndcounts,m,maxnz;
1469: PetscInt header[4],*rowlengths = 0,M,N,*cols;
1470: PetscInt *ourlens,*procsnz = 0,*offlens,jj,*mycols,*smycols;
1471: PetscInt i,nz,j,rstart,rend;
1472: int fd;
1476: MPI_Comm_size(comm,&size);
1477: MPI_Comm_rank(comm,&rank);
1478: if (!rank) {
1479: PetscViewerBinaryGetDescriptor(viewer,&fd);
1480: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
1481: if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
1482: }
1484: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
1485: M = header[1]; N = header[2]; nz = header[3];
1487: /*
1488: Handle case where matrix is stored on disk as a dense matrix
1489: */
1490: if (nz == MATRIX_BINARY_FORMAT_DENSE) {
1491: MatLoad_MPIDense_DenseInFile(comm,fd,M,N,type,newmat);
1492: return(0);
1493: }
1495: /* determine ownership of all rows */
1496: m = M/size + ((M % size) > rank);
1497: PetscMalloc((size+2)*sizeof(PetscInt),&rowners);
1498: MPI_Allgather(&m,1,MPI_INT,rowners+1,1,MPI_INT,comm);
1499: rowners[0] = 0;
1500: for (i=2; i<=size; i++) {
1501: rowners[i] += rowners[i-1];
1502: }
1503: rstart = rowners[rank];
1504: rend = rowners[rank+1];
1506: /* distribute row lengths to all processors */
1507: PetscMalloc(2*(rend-rstart+1)*sizeof(PetscInt),&ourlens);
1508: offlens = ourlens + (rend-rstart);
1509: if (!rank) {
1510: PetscMalloc(M*sizeof(PetscInt),&rowlengths);
1511: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
1512: PetscMalloc(size*sizeof(PetscMPIInt),&sndcounts);
1513: for (i=0; i<size; i++) sndcounts[i] = rowners[i+1] - rowners[i];
1514: MPI_Scatterv(rowlengths,sndcounts,rowners,MPIU_INT,ourlens,rend-rstart,MPIU_INT,0,comm);
1515: PetscFree(sndcounts);
1516: } else {
1517: MPI_Scatterv(0,0,0,MPIU_INT,ourlens,rend-rstart,MPIU_INT,0,comm);
1518: }
1520: if (!rank) {
1521: /* calculate the number of nonzeros on each processor */
1522: PetscMalloc(size*sizeof(PetscInt),&procsnz);
1523: PetscMemzero(procsnz,size*sizeof(PetscInt));
1524: for (i=0; i<size; i++) {
1525: for (j=rowners[i]; j< rowners[i+1]; j++) {
1526: procsnz[i] += rowlengths[j];
1527: }
1528: }
1529: PetscFree(rowlengths);
1531: /* determine max buffer needed and allocate it */
1532: maxnz = 0;
1533: for (i=0; i<size; i++) {
1534: maxnz = PetscMax(maxnz,procsnz[i]);
1535: }
1536: PetscMalloc(maxnz*sizeof(PetscInt),&cols);
1538: /* read in my part of the matrix column indices */
1539: nz = procsnz[0];
1540: PetscMalloc(nz*sizeof(PetscInt),&mycols);
1541: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
1543: /* read in every one elses and ship off */
1544: for (i=1; i<size; i++) {
1545: nz = procsnz[i];
1546: PetscBinaryRead(fd,cols,nz,PETSC_INT);
1547: MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
1548: }
1549: PetscFree(cols);
1550: } else {
1551: /* determine buffer space needed for message */
1552: nz = 0;
1553: for (i=0; i<m; i++) {
1554: nz += ourlens[i];
1555: }
1556: PetscMalloc((nz+1)*sizeof(PetscInt),&mycols);
1558: /* receive message of column indices*/
1559: MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
1560: MPI_Get_count(&status,MPIU_INT,&maxnz);
1561: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
1562: }
1564: /* loop over local rows, determining number of off diagonal entries */
1565: PetscMemzero(offlens,m*sizeof(PetscInt));
1566: jj = 0;
1567: for (i=0; i<m; i++) {
1568: for (j=0; j<ourlens[i]; j++) {
1569: if (mycols[jj] < rstart || mycols[jj] >= rend) offlens[i]++;
1570: jj++;
1571: }
1572: }
1574: /* create our matrix */
1575: for (i=0; i<m; i++) {
1576: ourlens[i] -= offlens[i];
1577: }
1578: MatCreate(comm,newmat);
1579: MatSetSizes(*newmat,m,PETSC_DECIDE,M,N);
1580: MatSetType(*newmat,type);
1581: MatMPIDenseSetPreallocation(*newmat,PETSC_NULL);
1582: A = *newmat;
1583: for (i=0; i<m; i++) {
1584: ourlens[i] += offlens[i];
1585: }
1587: if (!rank) {
1588: PetscMalloc(maxnz*sizeof(PetscScalar),&vals);
1590: /* read in my part of the matrix numerical values */
1591: nz = procsnz[0];
1592: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1593:
1594: /* insert into matrix */
1595: jj = rstart;
1596: smycols = mycols;
1597: svals = vals;
1598: for (i=0; i<m; i++) {
1599: MatSetValues(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
1600: smycols += ourlens[i];
1601: svals += ourlens[i];
1602: jj++;
1603: }
1605: /* read in other processors and ship out */
1606: for (i=1; i<size; i++) {
1607: nz = procsnz[i];
1608: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1609: MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)A)->tag,comm);
1610: }
1611: PetscFree(procsnz);
1612: } else {
1613: /* receive numeric values */
1614: PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);
1616: /* receive message of values*/
1617: MPI_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)A)->tag,comm,&status);
1618: MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
1619: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
1621: /* insert into matrix */
1622: jj = rstart;
1623: smycols = mycols;
1624: svals = vals;
1625: for (i=0; i<m; i++) {
1626: MatSetValues(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
1627: smycols += ourlens[i];
1628: svals += ourlens[i];
1629: jj++;
1630: }
1631: }
1632: PetscFree(ourlens);
1633: PetscFree(vals);
1634: PetscFree(mycols);
1635: PetscFree(rowners);
1637: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1638: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1639: return(0);
1640: }
1644: PetscErrorCode MatEqual_MPIDense(Mat A,Mat B,PetscTruth *flag)
1645: {
1646: Mat_MPIDense *matB = (Mat_MPIDense*)B->data,*matA = (Mat_MPIDense*)A->data;
1647: Mat a,b;
1648: PetscTruth flg;
1652: a = matA->A;
1653: b = matB->A;
1654: MatEqual(a,b,&flg);
1655: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);
1656: return(0);
1657: }