Actual source code: matelem.cxx
1: #include <petsc/private/petscelemental.h>
3: const char ElementalCitation[] = "@Article{Elemental2012,\n"
4: " author = {Jack Poulson and Bryan Marker and Jeff R. Hammond and Nichols A. Romero and Robert {v}an~{d}e~{G}eijn},\n"
5: " title = {Elemental: A New Framework for Distributed Memory Dense Matrix Computations},\n"
6: " journal = {{ACM} Transactions on Mathematical Software},\n"
7: " volume = {39},\n"
8: " number = {2},\n"
9: " year = {2013}\n"
10: "}\n";
11: static PetscBool ElementalCite = PETSC_FALSE;
13: /*
14: The variable Petsc_Elemental_keyval is used to indicate an MPI attribute that
15: is attached to a communicator, in this case the attribute is a Mat_Elemental_Grid
16: */
17: static PetscMPIInt Petsc_Elemental_keyval = MPI_KEYVAL_INVALID;
19: static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
20: {
22: Mat_Elemental *a = (Mat_Elemental*)A->data;
23: PetscBool iascii;
26: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
27: if (iascii) {
28: PetscViewerFormat format;
29: PetscViewerGetFormat(viewer,&format);
30: if (format == PETSC_VIEWER_ASCII_INFO) {
31: /* call elemental viewing function */
32: PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");
33: PetscViewerASCIIPrintf(viewer," allocated entries=%d\n",(*a->emat).AllocatedMemory());
34: PetscViewerASCIIPrintf(viewer," grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());
35: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
36: /* call elemental viewing function */
37: PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");
38: }
40: } else if (format == PETSC_VIEWER_DEFAULT) {
41: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
42: El::Print( *a->emat, "Elemental matrix (cyclic ordering)");
43: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
44: if (A->factortype == MAT_FACTOR_NONE){
45: Mat Adense;
46: MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);
47: MatView(Adense,viewer);
48: MatDestroy(&Adense);
49: }
50: } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
51: } else {
52: /* convert to dense format and call MatView() */
53: Mat Adense;
54: MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);
55: MatView(Adense,viewer);
56: MatDestroy(&Adense);
57: }
58: return(0);
59: }
61: static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
62: {
63: Mat_Elemental *a = (Mat_Elemental*)A->data;
66: info->block_size = 1.0;
68: if (flag == MAT_LOCAL) {
69: info->nz_allocated = (*a->emat).AllocatedMemory(); /* locally allocated */
70: info->nz_used = info->nz_allocated;
71: } else if (flag == MAT_GLOBAL_MAX) {
72: //MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));
73: /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
74: //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
75: } else if (flag == MAT_GLOBAL_SUM) {
76: //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
77: info->nz_allocated = (*a->emat).AllocatedMemory(); /* locally allocated */
78: info->nz_used = info->nz_allocated; /* assume Elemental does accurate allocation */
79: //MPIU_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
80: //PetscPrintf(PETSC_COMM_SELF," ... [%d] locally allocated %g\n",rank,info->nz_allocated);
81: }
83: info->nz_unneeded = 0.0;
84: info->assemblies = A->num_ass;
85: info->mallocs = 0;
86: info->memory = ((PetscObject)A)->mem;
87: info->fill_ratio_given = 0; /* determined by Elemental */
88: info->fill_ratio_needed = 0;
89: info->factor_mallocs = 0;
90: return(0);
91: }
93: PetscErrorCode MatSetOption_Elemental(Mat A,MatOption op,PetscBool flg)
94: {
95: Mat_Elemental *a = (Mat_Elemental*)A->data;
98: switch (op) {
99: case MAT_NEW_NONZERO_LOCATIONS:
100: case MAT_NEW_NONZERO_LOCATION_ERR:
101: case MAT_NEW_NONZERO_ALLOCATION_ERR:
102: case MAT_SYMMETRIC:
103: case MAT_SORTED_FULL:
104: case MAT_HERMITIAN:
105: break;
106: case MAT_ROW_ORIENTED:
107: a->roworiented = flg;
108: break;
109: default:
110: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %s",MatOptions[op]);
111: }
112: return(0);
113: }
115: static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode)
116: {
117: Mat_Elemental *a = (Mat_Elemental*)A->data;
118: PetscInt i,j,rrank,ridx,crank,cidx,erow,ecol,numQueues=0;
121: // TODO: Initialize matrix to all zeros?
123: // Count the number of queues from this process
124: if (a->roworiented) {
125: for (i=0; i<nr; i++) {
126: if (rows[i] < 0) continue;
127: P2RO(A,0,rows[i],&rrank,&ridx);
128: RO2E(A,0,rrank,ridx,&erow);
129: if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
130: for (j=0; j<nc; j++) {
131: if (cols[j] < 0) continue;
132: P2RO(A,1,cols[j],&crank,&cidx);
133: RO2E(A,1,crank,cidx,&ecol);
134: if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
135: if (!a->emat->IsLocal(erow,ecol)){ /* off-proc entry */
136: /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */
137: if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
138: ++numQueues;
139: continue;
140: }
141: /* printf("Locally updating (%d,%d)\n",erow,ecol); */
142: switch (imode) {
143: case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break;
144: case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break;
145: default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
146: }
147: }
148: }
150: /* printf("numQueues=%d\n",numQueues); */
151: a->emat->Reserve( numQueues);
152: for (i=0; i<nr; i++) {
153: if (rows[i] < 0) continue;
154: P2RO(A,0,rows[i],&rrank,&ridx);
155: RO2E(A,0,rrank,ridx,&erow);
156: for (j=0; j<nc; j++) {
157: if (cols[j] < 0) continue;
158: P2RO(A,1,cols[j],&crank,&cidx);
159: RO2E(A,1,crank,cidx,&ecol);
160: if (!a->emat->IsLocal(erow,ecol)) { /*off-proc entry*/
161: /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */
162: a->emat->QueueUpdate( erow, ecol, vals[i*nc+j]);
163: }
164: }
165: }
166: } else { /* columnoriented */
167: for (j=0; j<nc; j++) {
168: if (cols[j] < 0) continue;
169: P2RO(A,1,cols[j],&crank,&cidx);
170: RO2E(A,1,crank,cidx,&ecol);
171: if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
172: for (i=0; i<nr; i++) {
173: if (rows[i] < 0) continue;
174: P2RO(A,0,rows[i],&rrank,&ridx);
175: RO2E(A,0,rrank,ridx,&erow);
176: if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
177: if (!a->emat->IsLocal(erow,ecol)){ /* off-proc entry */
178: /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */
179: if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
180: ++numQueues;
181: continue;
182: }
183: /* printf("Locally updating (%d,%d)\n",erow,ecol); */
184: switch (imode) {
185: case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break;
186: case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break;
187: default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
188: }
189: }
190: }
192: /* printf("numQueues=%d\n",numQueues); */
193: a->emat->Reserve( numQueues);
194: for (j=0; j<nc; j++) {
195: if (cols[j] < 0) continue;
196: P2RO(A,1,cols[j],&crank,&cidx);
197: RO2E(A,1,crank,cidx,&ecol);
199: for (i=0; i<nr; i++) {
200: if (rows[i] < 0) continue;
201: P2RO(A,0,rows[i],&rrank,&ridx);
202: RO2E(A,0,rrank,ridx,&erow);
203: if (!a->emat->IsLocal(erow,ecol)) { /*off-proc entry*/
204: /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */
205: a->emat->QueueUpdate( erow, ecol, vals[i+j*nr]);
206: }
207: }
208: }
209: }
210: return(0);
211: }
213: static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
214: {
215: Mat_Elemental *a = (Mat_Elemental*)A->data;
216: PetscErrorCode ierr;
217: const PetscElemScalar *x;
218: PetscElemScalar *y;
219: PetscElemScalar one = 1,zero = 0;
222: VecGetArrayRead(X,(const PetscScalar **)&x);
223: VecGetArray(Y,(PetscScalar **)&y);
224: { /* Scoping so that constructor is called before pointer is returned */
225: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
226: xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
227: ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n);
228: El::Gemv(El::NORMAL,one,*a->emat,xe,zero,ye);
229: }
230: VecRestoreArrayRead(X,(const PetscScalar **)&x);
231: VecRestoreArray(Y,(PetscScalar **)&y);
232: return(0);
233: }
235: static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
236: {
237: Mat_Elemental *a = (Mat_Elemental*)A->data;
238: PetscErrorCode ierr;
239: const PetscElemScalar *x;
240: PetscElemScalar *y;
241: PetscElemScalar one = 1,zero = 0;
244: VecGetArrayRead(X,(const PetscScalar **)&x);
245: VecGetArray(Y,(PetscScalar **)&y);
246: { /* Scoping so that constructor is called before pointer is returned */
247: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
248: xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
249: ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n);
250: El::Gemv(El::TRANSPOSE,one,*a->emat,xe,zero,ye);
251: }
252: VecRestoreArrayRead(X,(const PetscScalar **)&x);
253: VecRestoreArray(Y,(PetscScalar **)&y);
254: return(0);
255: }
257: static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
258: {
259: Mat_Elemental *a = (Mat_Elemental*)A->data;
260: PetscErrorCode ierr;
261: const PetscElemScalar *x;
262: PetscElemScalar *z;
263: PetscElemScalar one = 1;
266: if (Y != Z) {VecCopy(Y,Z);}
267: VecGetArrayRead(X,(const PetscScalar **)&x);
268: VecGetArray(Z,(PetscScalar **)&z);
269: { /* Scoping so that constructor is called before pointer is returned */
270: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
271: xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
272: ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n);
273: El::Gemv(El::NORMAL,one,*a->emat,xe,one,ze);
274: }
275: VecRestoreArrayRead(X,(const PetscScalar **)&x);
276: VecRestoreArray(Z,(PetscScalar **)&z);
277: return(0);
278: }
280: static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
281: {
282: Mat_Elemental *a = (Mat_Elemental*)A->data;
283: PetscErrorCode ierr;
284: const PetscElemScalar *x;
285: PetscElemScalar *z;
286: PetscElemScalar one = 1;
289: if (Y != Z) {VecCopy(Y,Z);}
290: VecGetArrayRead(X,(const PetscScalar **)&x);
291: VecGetArray(Z,(PetscScalar **)&z);
292: { /* Scoping so that constructor is called before pointer is returned */
293: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
294: xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
295: ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n);
296: El::Gemv(El::TRANSPOSE,one,*a->emat,xe,one,ze);
297: }
298: VecRestoreArrayRead(X,(const PetscScalar **)&x);
299: VecRestoreArray(Z,(PetscScalar **)&z);
300: return(0);
301: }
303: PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
304: {
305: Mat_Elemental *a = (Mat_Elemental*)A->data;
306: Mat_Elemental *b = (Mat_Elemental*)B->data;
307: Mat_Elemental *c = (Mat_Elemental*)C->data;
308: PetscElemScalar one = 1,zero = 0;
311: { /* Scoping so that constructor is called before pointer is returned */
312: El::Gemm(El::NORMAL,El::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
313: }
314: C->assembled = PETSC_TRUE;
315: return(0);
316: }
318: PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat Ce)
319: {
323: MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);
324: MatSetType(Ce,MATELEMENTAL);
325: MatSetUp(Ce);
326: Ce->ops->matmultnumeric = MatMatMultNumeric_Elemental;
327: return(0);
328: }
330: static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
331: {
332: Mat_Elemental *a = (Mat_Elemental*)A->data;
333: Mat_Elemental *b = (Mat_Elemental*)B->data;
334: Mat_Elemental *c = (Mat_Elemental*)C->data;
335: PetscElemScalar one = 1,zero = 0;
338: { /* Scoping so that constructor is called before pointer is returned */
339: El::Gemm(El::NORMAL,El::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
340: }
341: C->assembled = PETSC_TRUE;
342: return(0);
343: }
345: static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat C)
346: {
350: MatSetSizes(C,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);
351: MatSetType(C,MATELEMENTAL);
352: MatSetUp(C);
353: return(0);
354: }
356: /* --------------------------------------- */
357: static PetscErrorCode MatProductSetFromOptions_Elemental_AB(Mat C)
358: {
360: C->ops->matmultsymbolic = MatMatMultSymbolic_Elemental;
361: C->ops->productsymbolic = MatProductSymbolic_AB;
362: return(0);
363: }
365: static PetscErrorCode MatProductSetFromOptions_Elemental_ABt(Mat C)
366: {
368: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_Elemental;
369: C->ops->productsymbolic = MatProductSymbolic_ABt;
370: return(0);
371: }
373: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Elemental(Mat C)
374: {
376: Mat_Product *product = C->product;
379: switch (product->type) {
380: case MATPRODUCT_AB:
381: MatProductSetFromOptions_Elemental_AB(C);
382: break;
383: case MATPRODUCT_ABt:
384: MatProductSetFromOptions_Elemental_ABt(C);
385: break;
386: default:
387: break;
388: }
389: return(0);
390: }
392: PetscErrorCode MatMatMultNumeric_Elemental_MPIDense(Mat A,Mat B,Mat C)
393: {
394: Mat Be,Ce;
398: MatConvert(B,MATELEMENTAL,MAT_INITIAL_MATRIX,&Be);
399: MatMatMult(A,Be,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&Ce);
400: MatConvert(Ce,MATMPIDENSE,MAT_REUSE_MATRIX,&C);
401: MatDestroy(&Be);
402: MatDestroy(&Ce);
403: return(0);
404: }
406: PetscErrorCode MatMatMultSymbolic_Elemental_MPIDense(Mat A,Mat B,PetscReal fill,Mat C)
407: {
411: MatSetSizes(C,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);
412: MatSetType(C,MATMPIDENSE);
413: MatSetUp(C);
414: C->ops->matmultnumeric = MatMatMultNumeric_Elemental_MPIDense;
415: return(0);
416: }
418: PetscErrorCode MatProductSetFromOptions_Elemental_MPIDense_AB(Mat C)
419: {
421: C->ops->matmultsymbolic = MatMatMultSymbolic_Elemental_MPIDense;
422: C->ops->productsymbolic = MatProductSymbolic_AB;
423: return(0);
424: }
426: PetscErrorCode MatProductSetFromOptions_Elemental_MPIDense(Mat C)
427: {
429: Mat_Product *product = C->product;
432: if (product->type == MATPRODUCT_AB) {
433: MatProductSetFromOptions_Elemental_MPIDense_AB(C);
434: }
435: return(0);
436: }
437: /* --------------------------------------- */
439: static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
440: {
441: PetscInt i,nrows,ncols,nD,rrank,ridx,crank,cidx;
442: Mat_Elemental *a = (Mat_Elemental*)A->data;
443: PetscErrorCode ierr;
444: PetscElemScalar v;
445: MPI_Comm comm;
448: PetscObjectGetComm((PetscObject)A,&comm);
449: MatGetSize(A,&nrows,&ncols);
450: nD = nrows>ncols ? ncols : nrows;
451: for (i=0; i<nD; i++) {
452: PetscInt erow,ecol;
453: P2RO(A,0,i,&rrank,&ridx);
454: RO2E(A,0,rrank,ridx,&erow);
455: if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
456: P2RO(A,1,i,&crank,&cidx);
457: RO2E(A,1,crank,cidx,&ecol);
458: if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
459: v = a->emat->Get(erow,ecol);
460: VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);
461: }
462: VecAssemblyBegin(D);
463: VecAssemblyEnd(D);
464: return(0);
465: }
467: static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
468: {
469: Mat_Elemental *x = (Mat_Elemental*)X->data;
470: const PetscElemScalar *d;
471: PetscErrorCode ierr;
474: if (R) {
475: VecGetArrayRead(R,(const PetscScalar **)&d);
476: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
477: de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n);
478: El::DiagonalScale(El::RIGHT,El::NORMAL,de,*x->emat);
479: VecRestoreArrayRead(R,(const PetscScalar **)&d);
480: }
481: if (L) {
482: VecGetArrayRead(L,(const PetscScalar **)&d);
483: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
484: de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n);
485: El::DiagonalScale(El::LEFT,El::NORMAL,de,*x->emat);
486: VecRestoreArrayRead(L,(const PetscScalar **)&d);
487: }
488: return(0);
489: }
491: static PetscErrorCode MatMissingDiagonal_Elemental(Mat A,PetscBool *missing,PetscInt *d)
492: {
494: *missing = PETSC_FALSE;
495: return(0);
496: }
498: static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a)
499: {
500: Mat_Elemental *x = (Mat_Elemental*)X->data;
503: El::Scale((PetscElemScalar)a,*x->emat);
504: return(0);
505: }
507: /*
508: MatAXPY - Computes Y = a*X + Y.
509: */
510: static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
511: {
512: Mat_Elemental *x = (Mat_Elemental*)X->data;
513: Mat_Elemental *y = (Mat_Elemental*)Y->data;
517: El::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
518: PetscObjectStateIncrease((PetscObject)Y);
519: return(0);
520: }
522: static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
523: {
524: Mat_Elemental *a=(Mat_Elemental*)A->data;
525: Mat_Elemental *b=(Mat_Elemental*)B->data;
529: El::Copy(*a->emat,*b->emat);
530: PetscObjectStateIncrease((PetscObject)B);
531: return(0);
532: }
534: static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
535: {
536: Mat Be;
537: MPI_Comm comm;
538: Mat_Elemental *a=(Mat_Elemental*)A->data;
542: PetscObjectGetComm((PetscObject)A,&comm);
543: MatCreate(comm,&Be);
544: MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);
545: MatSetType(Be,MATELEMENTAL);
546: MatSetUp(Be);
547: *B = Be;
548: if (op == MAT_COPY_VALUES) {
549: Mat_Elemental *b=(Mat_Elemental*)Be->data;
550: El::Copy(*a->emat,*b->emat);
551: }
552: Be->assembled = PETSC_TRUE;
553: return(0);
554: }
556: static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
557: {
558: Mat Be = *B;
560: MPI_Comm comm;
561: Mat_Elemental *a = (Mat_Elemental*)A->data, *b;
564: PetscObjectGetComm((PetscObject)A,&comm);
565: /* Only out-of-place supported */
566: if (reuse == MAT_INPLACE_MATRIX) SETERRQ(comm,PETSC_ERR_SUP,"Only out-of-place supported");
567: if (reuse == MAT_INITIAL_MATRIX) {
568: MatCreate(comm,&Be);
569: MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);
570: MatSetType(Be,MATELEMENTAL);
571: MatSetUp(Be);
572: *B = Be;
573: }
574: b = (Mat_Elemental*)Be->data;
575: El::Transpose(*a->emat,*b->emat);
576: Be->assembled = PETSC_TRUE;
577: return(0);
578: }
580: static PetscErrorCode MatConjugate_Elemental(Mat A)
581: {
582: Mat_Elemental *a = (Mat_Elemental*)A->data;
585: El::Conjugate(*a->emat);
586: return(0);
587: }
589: static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
590: {
591: Mat Be = *B;
593: MPI_Comm comm;
594: Mat_Elemental *a = (Mat_Elemental*)A->data, *b;
597: PetscObjectGetComm((PetscObject)A,&comm);
598: /* Only out-of-place supported */
599: if (reuse == MAT_INITIAL_MATRIX){
600: MatCreate(comm,&Be);
601: MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);
602: MatSetType(Be,MATELEMENTAL);
603: MatSetUp(Be);
604: *B = Be;
605: }
606: b = (Mat_Elemental*)Be->data;
607: El::Adjoint(*a->emat,*b->emat);
608: Be->assembled = PETSC_TRUE;
609: return(0);
610: }
612: static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
613: {
614: Mat_Elemental *a = (Mat_Elemental*)A->data;
615: PetscErrorCode ierr;
616: PetscElemScalar *x;
617: PetscInt pivoting = a->pivoting;
620: VecCopy(B,X);
621: VecGetArray(X,(PetscScalar **)&x);
623: El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe;
624: xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
625: El::DistMatrix<PetscElemScalar,El::MC,El::MR> xer(xe);
626: switch (A->factortype) {
627: case MAT_FACTOR_LU:
628: if (pivoting == 0) {
629: El::lu::SolveAfter(El::NORMAL,*a->emat,xer);
630: } else if (pivoting == 1) {
631: El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,xer);
632: } else { /* pivoting == 2 */
633: El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,xer);
634: }
635: break;
636: case MAT_FACTOR_CHOLESKY:
637: El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,xer);
638: break;
639: default:
640: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
641: break;
642: }
643: El::Copy(xer,xe);
645: VecRestoreArray(X,(PetscScalar **)&x);
646: return(0);
647: }
649: static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
650: {
651: PetscErrorCode ierr;
654: MatSolve_Elemental(A,B,X);
655: VecAXPY(X,1,Y);
656: return(0);
657: }
659: static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
660: {
661: Mat_Elemental *a = (Mat_Elemental*)A->data;
662: Mat_Elemental *x;
663: Mat C;
664: PetscInt pivoting = a->pivoting;
665: PetscBool flg;
666: MatType type;
670: MatGetType(X,&type);
671: PetscStrcmp(type,MATELEMENTAL,&flg);
672: if (!flg) {
673: MatConvert(B,MATELEMENTAL,MAT_INITIAL_MATRIX,&C);
674: x = (Mat_Elemental*)C->data;
675: } else {
676: x = (Mat_Elemental*)X->data;
677: El::Copy(*((Mat_Elemental*)B->data)->emat,*x->emat);
678: }
679: switch (A->factortype) {
680: case MAT_FACTOR_LU:
681: if (pivoting == 0) {
682: El::lu::SolveAfter(El::NORMAL,*a->emat,*x->emat);
683: } else if (pivoting == 1) {
684: El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*x->emat);
685: } else {
686: El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,*x->emat);
687: }
688: break;
689: case MAT_FACTOR_CHOLESKY:
690: El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,*x->emat);
691: break;
692: default:
693: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
694: break;
695: }
696: if (!flg) {
697: MatConvert(C,type,MAT_REUSE_MATRIX,&X);
698: MatDestroy(&C);
699: }
700: return(0);
701: }
703: static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
704: {
705: Mat_Elemental *a = (Mat_Elemental*)A->data;
707: PetscInt pivoting = a->pivoting;
710: if (pivoting == 0) {
711: El::LU(*a->emat);
712: } else if (pivoting == 1) {
713: El::LU(*a->emat,*a->P);
714: } else {
715: El::LU(*a->emat,*a->P,*a->Q);
716: }
717: A->factortype = MAT_FACTOR_LU;
718: A->assembled = PETSC_TRUE;
720: PetscFree(A->solvertype);
721: PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);
722: return(0);
723: }
725: static PetscErrorCode MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
726: {
730: MatCopy(A,F,SAME_NONZERO_PATTERN);
731: MatLUFactor_Elemental(F,0,0,info);
732: return(0);
733: }
735: static PetscErrorCode MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
736: {
738: /* F is created and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
739: return(0);
740: }
742: static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
743: {
744: Mat_Elemental *a = (Mat_Elemental*)A->data;
745: El::DistMatrix<PetscElemScalar,El::MC,El::STAR> d;
749: El::Cholesky(El::UPPER,*a->emat);
750: A->factortype = MAT_FACTOR_CHOLESKY;
751: A->assembled = PETSC_TRUE;
753: PetscFree(A->solvertype);
754: PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);
755: return(0);
756: }
758: static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
759: {
763: MatCopy(A,F,SAME_NONZERO_PATTERN);
764: MatCholeskyFactor_Elemental(F,0,info);
765: return(0);
766: }
768: static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
769: {
771: /* F is created and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
772: return(0);
773: }
775: PetscErrorCode MatFactorGetSolverType_elemental_elemental(Mat A,MatSolverType *type)
776: {
778: *type = MATSOLVERELEMENTAL;
779: return(0);
780: }
782: static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
783: {
784: Mat B;
788: /* Create the factorization matrix */
789: MatCreate(PetscObjectComm((PetscObject)A),&B);
790: MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);
791: MatSetType(B,MATELEMENTAL);
792: MatSetUp(B);
793: B->factortype = ftype;
794: PetscFree(B->solvertype);
795: PetscStrallocpy(MATSOLVERELEMENTAL,&B->solvertype);
797: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_elemental_elemental);
798: *F = B;
799: return(0);
800: }
802: PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_Elemental(void)
803: {
807: MatSolverTypeRegister(MATSOLVERELEMENTAL,MATELEMENTAL, MAT_FACTOR_LU,MatGetFactor_elemental_elemental);
808: MatSolverTypeRegister(MATSOLVERELEMENTAL,MATELEMENTAL, MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);
809: return(0);
810: }
812: static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
813: {
814: Mat_Elemental *a=(Mat_Elemental*)A->data;
817: switch (type){
818: case NORM_1:
819: *nrm = El::OneNorm(*a->emat);
820: break;
821: case NORM_FROBENIUS:
822: *nrm = El::FrobeniusNorm(*a->emat);
823: break;
824: case NORM_INFINITY:
825: *nrm = El::InfinityNorm(*a->emat);
826: break;
827: default:
828: SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Unsupported norm type");
829: }
830: return(0);
831: }
833: static PetscErrorCode MatZeroEntries_Elemental(Mat A)
834: {
835: Mat_Elemental *a=(Mat_Elemental*)A->data;
838: El::Zero(*a->emat);
839: return(0);
840: }
842: static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
843: {
844: Mat_Elemental *a = (Mat_Elemental*)A->data;
846: PetscInt i,m,shift,stride,*idx;
849: if (rows) {
850: m = a->emat->LocalHeight();
851: shift = a->emat->ColShift();
852: stride = a->emat->ColStride();
853: PetscMalloc1(m,&idx);
854: for (i=0; i<m; i++) {
855: PetscInt rank,offset;
856: E2RO(A,0,shift+i*stride,&rank,&offset);
857: RO2P(A,0,rank,offset,&idx[i]);
858: }
859: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);
860: }
861: if (cols) {
862: m = a->emat->LocalWidth();
863: shift = a->emat->RowShift();
864: stride = a->emat->RowStride();
865: PetscMalloc1(m,&idx);
866: for (i=0; i<m; i++) {
867: PetscInt rank,offset;
868: E2RO(A,1,shift+i*stride,&rank,&offset);
869: RO2P(A,1,rank,offset,&idx[i]);
870: }
871: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);
872: }
873: return(0);
874: }
876: static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
877: {
878: Mat Bmpi;
879: Mat_Elemental *a = (Mat_Elemental*)A->data;
880: MPI_Comm comm;
881: PetscErrorCode ierr;
882: IS isrows,iscols;
883: PetscInt rrank,ridx,crank,cidx,nrows,ncols,i,j,erow,ecol,elrow,elcol;
884: const PetscInt *rows,*cols;
885: PetscElemScalar v;
886: const El::Grid &grid = a->emat->Grid();
889: PetscObjectGetComm((PetscObject)A,&comm);
891: if (reuse == MAT_REUSE_MATRIX) {
892: Bmpi = *B;
893: } else {
894: MatCreate(comm,&Bmpi);
895: MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);
896: MatSetType(Bmpi,MATDENSE);
897: MatSetUp(Bmpi);
898: }
900: /* Get local entries of A */
901: MatGetOwnershipIS(A,&isrows,&iscols);
902: ISGetLocalSize(isrows,&nrows);
903: ISGetIndices(isrows,&rows);
904: ISGetLocalSize(iscols,&ncols);
905: ISGetIndices(iscols,&cols);
907: if (a->roworiented) {
908: for (i=0; i<nrows; i++) {
909: P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */
910: RO2E(A,0,rrank,ridx,&erow);
911: if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
912: for (j=0; j<ncols; j++) {
913: P2RO(A,1,cols[j],&crank,&cidx);
914: RO2E(A,1,crank,cidx,&ecol);
915: if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
917: elrow = erow / grid.MCSize(); /* Elemental local row index */
918: elcol = ecol / grid.MRSize(); /* Elemental local column index */
919: v = a->emat->GetLocal(elrow,elcol);
920: MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);
921: }
922: }
923: } else { /* column-oriented */
924: for (j=0; j<ncols; j++) {
925: P2RO(A,1,cols[j],&crank,&cidx);
926: RO2E(A,1,crank,cidx,&ecol);
927: if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
928: for (i=0; i<nrows; i++) {
929: P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */
930: RO2E(A,0,rrank,ridx,&erow);
931: if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
933: elrow = erow / grid.MCSize(); /* Elemental local row index */
934: elcol = ecol / grid.MRSize(); /* Elemental local column index */
935: v = a->emat->GetLocal(elrow,elcol);
936: MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);
937: }
938: }
939: }
940: MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);
941: MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);
942: if (reuse == MAT_INPLACE_MATRIX) {
943: MatHeaderReplace(A,&Bmpi);
944: } else {
945: *B = Bmpi;
946: }
947: ISDestroy(&isrows);
948: ISDestroy(&iscols);
949: return(0);
950: }
952: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
953: {
954: Mat mat_elemental;
955: PetscErrorCode ierr;
956: PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols;
957: const PetscInt *cols;
958: const PetscScalar *vals;
961: if (reuse == MAT_REUSE_MATRIX) {
962: mat_elemental = *newmat;
963: MatZeroEntries(mat_elemental);
964: } else {
965: MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);
966: MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);
967: MatSetType(mat_elemental,MATELEMENTAL);
968: MatSetUp(mat_elemental);
969: }
970: for (row=0; row<M; row++) {
971: MatGetRow(A,row,&ncols,&cols,&vals);
972: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
973: MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);
974: MatRestoreRow(A,row,&ncols,&cols,&vals);
975: }
976: MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);
977: MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);
979: if (reuse == MAT_INPLACE_MATRIX) {
980: MatHeaderReplace(A,&mat_elemental);
981: } else {
982: *newmat = mat_elemental;
983: }
984: return(0);
985: }
987: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
988: {
989: Mat mat_elemental;
990: PetscErrorCode ierr;
991: PetscInt row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j;
992: const PetscInt *cols;
993: const PetscScalar *vals;
996: if (reuse == MAT_REUSE_MATRIX) {
997: mat_elemental = *newmat;
998: MatZeroEntries(mat_elemental);
999: } else {
1000: MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);
1001: MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);
1002: MatSetType(mat_elemental,MATELEMENTAL);
1003: MatSetUp(mat_elemental);
1004: }
1005: for (row=rstart; row<rend; row++) {
1006: MatGetRow(A,row,&ncols,&cols,&vals);
1007: for (j=0; j<ncols; j++) {
1008: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1009: MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);
1010: }
1011: MatRestoreRow(A,row,&ncols,&cols,&vals);
1012: }
1013: MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);
1014: MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);
1016: if (reuse == MAT_INPLACE_MATRIX) {
1017: MatHeaderReplace(A,&mat_elemental);
1018: } else {
1019: *newmat = mat_elemental;
1020: }
1021: return(0);
1022: }
1024: PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
1025: {
1026: Mat mat_elemental;
1027: PetscErrorCode ierr;
1028: PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols,j;
1029: const PetscInt *cols;
1030: const PetscScalar *vals;
1033: if (reuse == MAT_REUSE_MATRIX) {
1034: mat_elemental = *newmat;
1035: MatZeroEntries(mat_elemental);
1036: } else {
1037: MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);
1038: MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);
1039: MatSetType(mat_elemental,MATELEMENTAL);
1040: MatSetUp(mat_elemental);
1041: }
1042: MatGetRowUpperTriangular(A);
1043: for (row=0; row<M; row++) {
1044: MatGetRow(A,row,&ncols,&cols,&vals);
1045: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1046: MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);
1047: for (j=0; j<ncols; j++) { /* lower triangular part */
1048: PetscScalar v;
1049: if (cols[j] == row) continue;
1050: v = A->hermitian ? PetscConj(vals[j]) : vals[j];
1051: MatSetValues(mat_elemental,1,&cols[j],1,&row,&v,ADD_VALUES);
1052: }
1053: MatRestoreRow(A,row,&ncols,&cols,&vals);
1054: }
1055: MatRestoreRowUpperTriangular(A);
1056: MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);
1057: MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);
1059: if (reuse == MAT_INPLACE_MATRIX) {
1060: MatHeaderReplace(A,&mat_elemental);
1061: } else {
1062: *newmat = mat_elemental;
1063: }
1064: return(0);
1065: }
1067: PETSC_INTERN PetscErrorCode MatConvert_MPISBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
1068: {
1069: Mat mat_elemental;
1070: PetscErrorCode ierr;
1071: PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols,j,rstart=A->rmap->rstart,rend=A->rmap->rend;
1072: const PetscInt *cols;
1073: const PetscScalar *vals;
1076: if (reuse == MAT_REUSE_MATRIX) {
1077: mat_elemental = *newmat;
1078: MatZeroEntries(mat_elemental);
1079: } else {
1080: MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);
1081: MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);
1082: MatSetType(mat_elemental,MATELEMENTAL);
1083: MatSetUp(mat_elemental);
1084: }
1085: MatGetRowUpperTriangular(A);
1086: for (row=rstart; row<rend; row++) {
1087: MatGetRow(A,row,&ncols,&cols,&vals);
1088: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1089: MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);
1090: for (j=0; j<ncols; j++) { /* lower triangular part */
1091: PetscScalar v;
1092: if (cols[j] == row) continue;
1093: v = A->hermitian ? PetscConj(vals[j]) : vals[j];
1094: MatSetValues(mat_elemental,1,&cols[j],1,&row,&v,ADD_VALUES);
1095: }
1096: MatRestoreRow(A,row,&ncols,&cols,&vals);
1097: }
1098: MatRestoreRowUpperTriangular(A);
1099: MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);
1100: MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);
1102: if (reuse == MAT_INPLACE_MATRIX) {
1103: MatHeaderReplace(A,&mat_elemental);
1104: } else {
1105: *newmat = mat_elemental;
1106: }
1107: return(0);
1108: }
1110: static PetscErrorCode MatDestroy_Elemental(Mat A)
1111: {
1112: Mat_Elemental *a = (Mat_Elemental*)A->data;
1113: PetscErrorCode ierr;
1114: Mat_Elemental_Grid *commgrid;
1115: PetscBool flg;
1116: MPI_Comm icomm;
1119: delete a->emat;
1120: delete a->P;
1121: delete a->Q;
1123: El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
1124: PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);
1125: MPI_Comm_get_attr(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);
1126: if (--commgrid->grid_refct == 0) {
1127: delete commgrid->grid;
1128: PetscFree(commgrid);
1129: MPI_Comm_free_keyval(&Petsc_Elemental_keyval);
1130: }
1131: PetscCommDestroy(&icomm);
1132: PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);
1133: PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);
1134: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_elemental_mpidense_C",NULL);
1135: PetscFree(A->data);
1136: return(0);
1137: }
1139: PetscErrorCode MatSetUp_Elemental(Mat A)
1140: {
1141: Mat_Elemental *a = (Mat_Elemental*)A->data;
1143: MPI_Comm comm;
1144: PetscMPIInt rsize,csize;
1145: PetscInt n;
1148: PetscLayoutSetUp(A->rmap);
1149: PetscLayoutSetUp(A->cmap);
1151: /* Check if local row and column sizes are equally distributed.
1152: Jed: Elemental uses "element" cyclic ordering so the sizes need to match that
1153: exactly. The strategy in MatElemental is for PETSc to implicitly permute to block ordering (like would be returned by
1154: PetscSplitOwnership(comm,&n,&N), at which point Elemental matrices can act on PETSc vectors without redistributing the vectors. */
1155: PetscObjectGetComm((PetscObject)A,&comm);
1156: n = PETSC_DECIDE;
1157: PetscSplitOwnership(comm,&n,&A->rmap->N);
1158: if (n != A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Local row size %D of ELEMENTAL matrix must be equally distributed",A->rmap->n);
1160: n = PETSC_DECIDE;
1161: PetscSplitOwnership(comm,&n,&A->cmap->N);
1162: if (n != A->cmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Local column size %D of ELEMENTAL matrix must be equally distributed",A->cmap->n);
1164: a->emat->Resize(A->rmap->N,A->cmap->N);
1165: El::Zero(*a->emat);
1167: MPI_Comm_size(A->rmap->comm,&rsize);
1168: MPI_Comm_size(A->cmap->comm,&csize);
1169: if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
1170: a->commsize = rsize;
1171: a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
1172: a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
1173: a->m[0] = A->rmap->N / rsize + (a->mr[0] != rsize);
1174: a->m[1] = A->cmap->N / csize + (a->mr[1] != csize);
1175: return(0);
1176: }
1178: PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
1179: {
1180: Mat_Elemental *a = (Mat_Elemental*)A->data;
1183: /* printf("Calling ProcessQueues\n"); */
1184: a->emat->ProcessQueues();
1185: /* printf("Finished ProcessQueues\n"); */
1186: return(0);
1187: }
1189: PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
1190: {
1192: /* Currently does nothing */
1193: return(0);
1194: }
1196: PetscErrorCode MatLoad_Elemental(Mat newMat, PetscViewer viewer)
1197: {
1199: Mat Adense,Ae;
1200: MPI_Comm comm;
1203: PetscObjectGetComm((PetscObject)newMat,&comm);
1204: MatCreate(comm,&Adense);
1205: MatSetType(Adense,MATDENSE);
1206: MatLoad(Adense,viewer);
1207: MatConvert(Adense, MATELEMENTAL, MAT_INITIAL_MATRIX,&Ae);
1208: MatDestroy(&Adense);
1209: MatHeaderReplace(newMat,&Ae);
1210: return(0);
1211: }
1213: /* -------------------------------------------------------------------*/
1214: static struct _MatOps MatOps_Values = {
1215: MatSetValues_Elemental,
1216: 0,
1217: 0,
1218: MatMult_Elemental,
1219: /* 4*/ MatMultAdd_Elemental,
1220: MatMultTranspose_Elemental,
1221: MatMultTransposeAdd_Elemental,
1222: MatSolve_Elemental,
1223: MatSolveAdd_Elemental,
1224: 0,
1225: /*10*/ 0,
1226: MatLUFactor_Elemental,
1227: MatCholeskyFactor_Elemental,
1228: 0,
1229: MatTranspose_Elemental,
1230: /*15*/ MatGetInfo_Elemental,
1231: 0,
1232: MatGetDiagonal_Elemental,
1233: MatDiagonalScale_Elemental,
1234: MatNorm_Elemental,
1235: /*20*/ MatAssemblyBegin_Elemental,
1236: MatAssemblyEnd_Elemental,
1237: MatSetOption_Elemental,
1238: MatZeroEntries_Elemental,
1239: /*24*/ 0,
1240: MatLUFactorSymbolic_Elemental,
1241: MatLUFactorNumeric_Elemental,
1242: MatCholeskyFactorSymbolic_Elemental,
1243: MatCholeskyFactorNumeric_Elemental,
1244: /*29*/ MatSetUp_Elemental,
1245: 0,
1246: 0,
1247: 0,
1248: 0,
1249: /*34*/ MatDuplicate_Elemental,
1250: 0,
1251: 0,
1252: 0,
1253: 0,
1254: /*39*/ MatAXPY_Elemental,
1255: 0,
1256: 0,
1257: 0,
1258: MatCopy_Elemental,
1259: /*44*/ 0,
1260: MatScale_Elemental,
1261: MatShift_Basic,
1262: 0,
1263: 0,
1264: /*49*/ 0,
1265: 0,
1266: 0,
1267: 0,
1268: 0,
1269: /*54*/ 0,
1270: 0,
1271: 0,
1272: 0,
1273: 0,
1274: /*59*/ 0,
1275: MatDestroy_Elemental,
1276: MatView_Elemental,
1277: 0,
1278: 0,
1279: /*64*/ 0,
1280: 0,
1281: 0,
1282: 0,
1283: 0,
1284: /*69*/ 0,
1285: 0,
1286: MatConvert_Elemental_Dense,
1287: 0,
1288: 0,
1289: /*74*/ 0,
1290: 0,
1291: 0,
1292: 0,
1293: 0,
1294: /*79*/ 0,
1295: 0,
1296: 0,
1297: 0,
1298: MatLoad_Elemental,
1299: /*84*/ 0,
1300: 0,
1301: 0,
1302: 0,
1303: 0,
1304: /*89*/ 0,
1305: 0,
1306: MatMatMultNumeric_Elemental,
1307: 0,
1308: 0,
1309: /*94*/ 0,
1310: 0,
1311: 0,
1312: MatMatTransposeMultNumeric_Elemental,
1313: 0,
1314: /*99*/ MatProductSetFromOptions_Elemental,
1315: 0,
1316: 0,
1317: MatConjugate_Elemental,
1318: 0,
1319: /*104*/0,
1320: 0,
1321: 0,
1322: 0,
1323: 0,
1324: /*109*/MatMatSolve_Elemental,
1325: 0,
1326: 0,
1327: 0,
1328: MatMissingDiagonal_Elemental,
1329: /*114*/0,
1330: 0,
1331: 0,
1332: 0,
1333: 0,
1334: /*119*/0,
1335: MatHermitianTranspose_Elemental,
1336: 0,
1337: 0,
1338: 0,
1339: /*124*/0,
1340: 0,
1341: 0,
1342: 0,
1343: 0,
1344: /*129*/0,
1345: 0,
1346: 0,
1347: 0,
1348: 0,
1349: /*134*/0,
1350: 0,
1351: 0,
1352: 0,
1353: 0,
1354: 0,
1355: /*140*/0,
1356: 0,
1357: 0,
1358: 0,
1359: 0,
1360: /*145*/0,
1361: 0,
1362: 0
1363: };
1365: /*MC
1366: MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1368: Use ./configure --download-elemental to install PETSc to use Elemental
1370: Use -pc_type lu -pc_factor_mat_solver_type elemental to use this direct solver
1372: Options Database Keys:
1373: + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
1374: - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1376: Level: beginner
1378: .seealso: MATDENSE
1379: M*/
1381: PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1382: {
1383: Mat_Elemental *a;
1384: PetscErrorCode ierr;
1385: PetscBool flg,flg1;
1386: Mat_Elemental_Grid *commgrid;
1387: MPI_Comm icomm;
1388: PetscInt optv1;
1391: PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));
1392: A->insertmode = NOT_SET_VALUES;
1394: PetscNewLog(A,&a);
1395: A->data = (void*)a;
1397: /* Set up the elemental matrix */
1398: El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
1400: /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
1401: if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
1402: MPI_Comm_create_keyval(MPI_COMM_NULL_COPY_FN,MPI_COMM_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);
1403: PetscCitationsRegister(ElementalCitation,&ElementalCite);
1404: }
1405: PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);
1406: MPI_Comm_get_attr(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);
1407: if (!flg) {
1408: PetscNewLog(A,&commgrid);
1410: PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");
1411: /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until El::Grid() is called */
1412: PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",El::mpi::Size(cxxcomm),&optv1,&flg1);
1413: if (flg1) {
1414: if (El::mpi::Size(cxxcomm) % optv1) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)El::mpi::Size(cxxcomm));
1415: commgrid->grid = new El::Grid(cxxcomm,optv1); /* use user-provided grid height */
1416: } else {
1417: commgrid->grid = new El::Grid(cxxcomm); /* use Elemental default grid sizes */
1418: /* printf("new commgrid->grid = %p\n",commgrid->grid); -- memory leak revealed by valgrind? */
1419: }
1420: commgrid->grid_refct = 1;
1421: MPI_Comm_set_attr(icomm,Petsc_Elemental_keyval,(void*)commgrid);
1423: a->pivoting = 1;
1424: PetscOptionsInt("-mat_elemental_pivoting","Pivoting","None",a->pivoting,&a->pivoting,NULL);
1426: PetscOptionsEnd();
1427: } else {
1428: commgrid->grid_refct++;
1429: }
1430: PetscCommDestroy(&icomm);
1431: a->grid = commgrid->grid;
1432: a->emat = new El::DistMatrix<PetscElemScalar>(*a->grid);
1433: a->roworiented = PETSC_TRUE;
1435: PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);
1436: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_elemental_mpidense_C",MatProductSetFromOptions_Elemental_MPIDense);
1437: PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);
1438: return(0);
1439: }