xref: /petsc/src/mat/impls/elemental/matelem.cxx (revision 0f39cd5a0d5f57ab1a62c821a7af94f4edff100f)
1 #include <../src/mat/impls/elemental/matelemimpl.h> /*I "petscmat.h" I*/
2 
3 /*
4     The variable Petsc_Elemental_keyval is used to indicate an MPI attribute that
5   is attached to a communicator, in this case the attribute is a Mat_Elemental_Grid
6 */
7 static PetscMPIInt Petsc_Elemental_keyval = MPI_KEYVAL_INVALID;
8 
9 #undef __FUNCT__
10 #define __FUNCT__ "PetscElementalInitializePackage"
11 /*@C
12    PetscElementalInitializePackage - Initialize Elemental package
13 
14    Logically Collective
15 
16    Level: developer
17 
18 .seealso: MATELEMENTAL, PetscElementalFinalizePackage()
19 @*/
20 PetscErrorCode PetscElementalInitializePackage(void)
21 {
22   PetscErrorCode ierr;
23 
24   PetscFunctionBegin;
25   if (elem::Initialized()) PetscFunctionReturn(0);
26   { /* We have already initialized MPI, so this song and dance is just to pass these variables (which won't be used by Elemental) through the interface that needs references */
27     int zero = 0;
28     char **nothing = 0;
29     elem::Initialize(zero,nothing);
30   }
31   ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr);
32   PetscFunctionReturn(0);
33 }
34 
35 #undef __FUNCT__
36 #define __FUNCT__ "PetscElementalFinalizePackage"
37 /*@C
38    PetscElementalFinalizePackage - Finalize Elemental package
39 
40    Logically Collective
41 
42    Level: developer
43 
44 .seealso: MATELEMENTAL, PetscElementalInitializePackage()
45 @*/
46 PetscErrorCode PetscElementalFinalizePackage(void)
47 {
48 
49   PetscFunctionBegin;
50   elem::Finalize();
51   PetscFunctionReturn(0);
52 }
53 
54 #undef __FUNCT__
55 #define __FUNCT__ "MatView_Elemental"
56 static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
57 {
58   PetscErrorCode ierr;
59   Mat_Elemental  *a = (Mat_Elemental*)A->data;
60   PetscBool      iascii;
61 
62   PetscFunctionBegin;
63   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
64   if (iascii) {
65     PetscViewerFormat format;
66     ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
67     if (format == PETSC_VIEWER_ASCII_INFO) {
68       /* call elemental viewing function */
69       ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr);
70       ierr = PetscViewerASCIIPrintf(viewer,"  allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr);
71       ierr = PetscViewerASCIIPrintf(viewer,"  grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr);
72       if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
73         /* call elemental viewing function */
74         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr);
75       }
76 
77     } else if (format == PETSC_VIEWER_DEFAULT) {
78       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr);
79       elem::Print( *a->emat, "Elemental matrix (cyclic ordering)" );
80       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr);
81       if (A->factortype == MAT_FACTOR_NONE){
82         Mat Adense;
83         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
84         ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
85         ierr = MatView(Adense,viewer);CHKERRQ(ierr);
86         ierr = MatDestroy(&Adense);CHKERRQ(ierr);
87       }
88     } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
89   } else {
90     /* convert to dense format and call MatView() */
91     Mat Adense;
92     ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
93     ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
94     ierr = MatView(Adense,viewer);CHKERRQ(ierr);
95     ierr = MatDestroy(&Adense);CHKERRQ(ierr);
96   }
97   PetscFunctionReturn(0);
98 }
99 
100 #undef __FUNCT__
101 #define __FUNCT__ "MatGetInfo_Elemental"
102 static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
103 {
104   Mat_Elemental  *a = (Mat_Elemental*)A->data;
105   PetscMPIInt    rank;
106 
107   PetscFunctionBegin;
108   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
109 
110   /* if (!rank) printf("          .........MatGetInfo_Elemental ...\n"); */
111   info->block_size     = 1.0;
112 
113   if (flag == MAT_LOCAL) {
114     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
115     info->nz_used        = info->nz_allocated;
116   } else if (flag == MAT_GLOBAL_MAX) {
117     //ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr);
118     /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
119     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
120   } else if (flag == MAT_GLOBAL_SUM) {
121     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
122     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
123     info->nz_used        = info->nz_allocated; /* assume Elemental does accurate allocation */
124     //ierr = MPI_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr);
125     //PetscPrintf(PETSC_COMM_SELF,"    ... [%d] locally allocated %g\n",rank,info->nz_allocated);
126   }
127 
128   info->nz_unneeded       = 0.0;
129   info->assemblies        = (double)A->num_ass;
130   info->mallocs           = 0;
131   info->memory            = ((PetscObject)A)->mem;
132   info->fill_ratio_given  = 0; /* determined by Elemental */
133   info->fill_ratio_needed = 0;
134   info->factor_mallocs    = 0;
135   PetscFunctionReturn(0);
136 }
137 
138 #undef __FUNCT__
139 #define __FUNCT__ "MatSetValues_Elemental"
140 static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode)
141 {
142   PetscErrorCode ierr;
143   Mat_Elemental  *a = (Mat_Elemental*)A->data;
144   PetscMPIInt    rank;
145   PetscInt       i,j,rrank,ridx,crank,cidx;
146 
147   PetscFunctionBegin;
148   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr);
149 
150   const elem::Grid &grid = a->emat->Grid();
151   for (i=0; i<nr; i++) {
152     PetscInt erow,ecol,elrow,elcol;
153     if (rows[i] < 0) continue;
154     P2RO(A,0,rows[i],&rrank,&ridx);
155     RO2E(A,0,rrank,ridx,&erow);
156     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
157     for (j=0; j<nc; j++) {
158       if (cols[j] < 0) continue;
159       P2RO(A,1,cols[j],&crank,&cidx);
160       RO2E(A,1,crank,cidx,&ecol);
161       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
162       if (erow % grid.MCSize() != grid.MCRank() || ecol % grid.MRSize() != grid.MRRank()){ /* off-proc entry */
163         if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
164         /* PetscPrintf(PETSC_COMM_SELF,"[%D] add off-proc entry (%D,%D, %g) (%D %D)\n",rank,rows[i],cols[j],*(vals+i*nc),erow,ecol); */
165         a->esubmat->Set(0,0, (PetscElemScalar)vals[i*nc+j]);
166         a->interface->Axpy(1.0,*(a->esubmat),erow,ecol);
167         continue;
168       }
169       elrow = erow / grid.MCSize();
170       elcol = ecol / grid.MRSize();
171       switch (imode) {
172       case INSERT_VALUES: a->emat->SetLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
173       case ADD_VALUES: a->emat->UpdateLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
174       default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
175       }
176     }
177   }
178   PetscFunctionReturn(0);
179 }
180 
181 #undef __FUNCT__
182 #define __FUNCT__ "MatMult_Elemental"
183 static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
184 {
185   Mat_Elemental         *a = (Mat_Elemental*)A->data;
186   PetscErrorCode        ierr;
187   const PetscElemScalar *x;
188   PetscElemScalar       *y;
189   PetscElemScalar       one = 1,zero = 0;
190 
191   PetscFunctionBegin;
192   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
193   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
194   { /* Scoping so that constructor is called before pointer is returned */
195     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ye;
196     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
197     ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n);
198     elem::Gemv(elem::NORMAL,one,*a->emat,xe,zero,ye);
199   }
200   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
201   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
202   PetscFunctionReturn(0);
203 }
204 
205 #undef __FUNCT__
206 #define __FUNCT__ "MatMultTranspose_Elemental"
207 static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
208 {
209   Mat_Elemental         *a = (Mat_Elemental*)A->data;
210   PetscErrorCode        ierr;
211   const PetscElemScalar *x;
212   PetscElemScalar       *y;
213   PetscElemScalar       one = 1,zero = 0;
214 
215   PetscFunctionBegin;
216   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
217   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
218   { /* Scoping so that constructor is called before pointer is returned */
219     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ye;
220     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
221     ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n);
222     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,zero,ye);
223   }
224   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
225   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
226   PetscFunctionReturn(0);
227 }
228 
229 #undef __FUNCT__
230 #define __FUNCT__ "MatMultAdd_Elemental"
231 static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
232 {
233   Mat_Elemental         *a = (Mat_Elemental*)A->data;
234   PetscErrorCode        ierr;
235   const PetscElemScalar *x;
236   PetscElemScalar       *z;
237   PetscElemScalar       one = 1;
238 
239   PetscFunctionBegin;
240   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
241   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
242   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
243   { /* Scoping so that constructor is called before pointer is returned */
244     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ze;
245     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
246     ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n);
247     elem::Gemv(elem::NORMAL,one,*a->emat,xe,one,ze);
248   }
249   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
250   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
251   PetscFunctionReturn(0);
252 }
253 
254 #undef __FUNCT__
255 #define __FUNCT__ "MatMultTransposeAdd_Elemental"
256 static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
257 {
258   Mat_Elemental         *a = (Mat_Elemental*)A->data;
259   PetscErrorCode        ierr;
260   const PetscElemScalar *x;
261   PetscElemScalar       *z;
262   PetscElemScalar       one = 1;
263 
264   PetscFunctionBegin;
265   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
266   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
267   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
268   { /* Scoping so that constructor is called before pointer is returned */
269     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ze;
270     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
271     ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n);
272     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,one,ze);
273   }
274   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
275   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
276   PetscFunctionReturn(0);
277 }
278 
279 #undef __FUNCT__
280 #define __FUNCT__ "MatMatMultNumeric_Elemental"
281 static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
282 {
283   Mat_Elemental    *a = (Mat_Elemental*)A->data;
284   Mat_Elemental    *b = (Mat_Elemental*)B->data;
285   Mat_Elemental    *c = (Mat_Elemental*)C->data;
286   PetscElemScalar  one = 1,zero = 0;
287 
288   PetscFunctionBegin;
289   { /* Scoping so that constructor is called before pointer is returned */
290     elem::Gemm(elem::NORMAL,elem::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
291   }
292   C->assembled = PETSC_TRUE;
293   PetscFunctionReturn(0);
294 }
295 
296 #undef __FUNCT__
297 #define __FUNCT__ "MatMatMultSymbolic_Elemental"
298 static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
299 {
300   PetscErrorCode ierr;
301   Mat            Ce;
302   MPI_Comm       comm;
303 
304   PetscFunctionBegin;
305   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
306   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
307   ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
308   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
309   ierr = MatSetUp(Ce);CHKERRQ(ierr);
310   *C = Ce;
311   PetscFunctionReturn(0);
312 }
313 
314 #undef __FUNCT__
315 #define __FUNCT__ "MatMatMult_Elemental"
316 static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
317 {
318   PetscErrorCode ierr;
319 
320   PetscFunctionBegin;
321   if (scall == MAT_INITIAL_MATRIX){
322     ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
323     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
324     ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
325   }
326   ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
327   ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
328   ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
329   PetscFunctionReturn(0);
330 }
331 
332 #undef __FUNCT__
333 #define __FUNCT__ "MatMatTransposeMultNumeric_Elemental"
334 static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
335 {
336   Mat_Elemental      *a = (Mat_Elemental*)A->data;
337   Mat_Elemental      *b = (Mat_Elemental*)B->data;
338   Mat_Elemental      *c = (Mat_Elemental*)C->data;
339   PetscElemScalar    one = 1,zero = 0;
340 
341   PetscFunctionBegin;
342   { /* Scoping so that constructor is called before pointer is returned */
343     elem::Gemm(elem::NORMAL,elem::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
344   }
345   C->assembled = PETSC_TRUE;
346   PetscFunctionReturn(0);
347 }
348 
349 #undef __FUNCT__
350 #define __FUNCT__ "MatMatTransposeMultSymbolic_Elemental"
351 static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
352 {
353   PetscErrorCode ierr;
354   Mat            Ce;
355   MPI_Comm       comm;
356 
357   PetscFunctionBegin;
358   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
359   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
360   ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
361   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
362   ierr = MatSetUp(Ce);CHKERRQ(ierr);
363   *C = Ce;
364   PetscFunctionReturn(0);
365 }
366 
367 #undef __FUNCT__
368 #define __FUNCT__ "MatMatTransposeMult_Elemental"
369 static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
370 {
371   PetscErrorCode ierr;
372 
373   PetscFunctionBegin;
374   if (scall == MAT_INITIAL_MATRIX){
375     ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
376     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
377     ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
378   }
379   ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
380   ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
381   ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
382   PetscFunctionReturn(0);
383 }
384 
385 #undef __FUNCT__
386 #define __FUNCT__ "MatGetDiagonal_Elemental"
387 static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
388 {
389   PetscInt        i,nrows,ncols,nD,rrank,ridx,crank,cidx;
390   Mat_Elemental   *a = (Mat_Elemental*)A->data;
391   PetscErrorCode  ierr;
392   PetscElemScalar v;
393   MPI_Comm        comm;
394 
395   PetscFunctionBegin;
396   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
397   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
398   nD = nrows>ncols ? ncols : nrows;
399   for (i=0; i<nD; i++) {
400     PetscInt erow,ecol;
401     P2RO(A,0,i,&rrank,&ridx);
402     RO2E(A,0,rrank,ridx,&erow);
403     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
404     P2RO(A,1,i,&crank,&cidx);
405     RO2E(A,1,crank,cidx,&ecol);
406     if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
407     v = a->emat->Get(erow,ecol);
408     ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr);
409   }
410   ierr = VecAssemblyBegin(D);CHKERRQ(ierr);
411   ierr = VecAssemblyEnd(D);CHKERRQ(ierr);
412   PetscFunctionReturn(0);
413 }
414 
415 #undef __FUNCT__
416 #define __FUNCT__ "MatDiagonalScale_Elemental"
417 static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
418 {
419   Mat_Elemental         *x = (Mat_Elemental*)X->data;
420   const PetscElemScalar *d;
421   PetscErrorCode        ierr;
422 
423   PetscFunctionBegin;
424   if (R) {
425     ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
426     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de;
427     de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n);
428     elem::DiagonalScale(elem::RIGHT,elem::NORMAL,de,*x->emat);
429     ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
430   }
431   if (L) {
432     ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
433     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de;
434     de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n);
435     elem::DiagonalScale(elem::LEFT,elem::NORMAL,de,*x->emat);
436     ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
437   }
438   PetscFunctionReturn(0);
439 }
440 
441 #undef __FUNCT__
442 #define __FUNCT__ "MatScale_Elemental"
443 static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a)
444 {
445   Mat_Elemental  *x = (Mat_Elemental*)X->data;
446 
447   PetscFunctionBegin;
448   elem::Scale((PetscElemScalar)a,*x->emat);
449   PetscFunctionReturn(0);
450 }
451 
452 #undef __FUNCT__
453 #define __FUNCT__ "MatAXPY_Elemental"
454 static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
455 {
456   Mat_Elemental  *x = (Mat_Elemental*)X->data;
457   Mat_Elemental  *y = (Mat_Elemental*)Y->data;
458   PetscErrorCode ierr;
459 
460   PetscFunctionBegin;
461   elem::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
462   ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr);
463   PetscFunctionReturn(0);
464 }
465 
466 #undef __FUNCT__
467 #define __FUNCT__ "MatCopy_Elemental"
468 static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
469 {
470   Mat_Elemental *a=(Mat_Elemental*)A->data;
471   Mat_Elemental *b=(Mat_Elemental*)B->data;
472 
473   PetscFunctionBegin;
474   elem::Copy(*a->emat,*b->emat);
475   PetscFunctionReturn(0);
476 }
477 
478 #undef __FUNCT__
479 #define __FUNCT__ "MatDuplicate_Elemental"
480 static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
481 {
482   Mat            Be;
483   MPI_Comm       comm;
484   Mat_Elemental  *a=(Mat_Elemental*)A->data;
485   PetscErrorCode ierr;
486 
487   PetscFunctionBegin;
488   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
489   ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
490   ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
491   ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
492   ierr = MatSetUp(Be);CHKERRQ(ierr);
493   *B = Be;
494   if (op == MAT_COPY_VALUES) {
495     Mat_Elemental *b=(Mat_Elemental*)Be->data;
496     elem::Copy(*a->emat,*b->emat);
497   }
498   Be->assembled = PETSC_TRUE;
499   PetscFunctionReturn(0);
500 }
501 
502 #undef __FUNCT__
503 #define __FUNCT__ "MatTranspose_Elemental"
504 static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
505 {
506   Mat            Be = *B;
507   PetscErrorCode ierr;
508   MPI_Comm       comm;
509   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
510 
511   PetscFunctionBegin;
512   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
513   /* Only out-of-place supported */
514   if (reuse == MAT_INITIAL_MATRIX){
515     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
516     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
517     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
518     ierr = MatSetUp(Be);CHKERRQ(ierr);
519     *B = Be;
520   }
521   b = (Mat_Elemental*)Be->data;
522   elem::Transpose(*a->emat,*b->emat);
523   Be->assembled = PETSC_TRUE;
524   PetscFunctionReturn(0);
525 }
526 
527 #undef __FUNCT__
528 #define __FUNCT__ "MatConjugate_Elemental"
529 static PetscErrorCode MatConjugate_Elemental(Mat A)
530 {
531   Mat_Elemental  *a = (Mat_Elemental*)A->data;
532 
533   PetscFunctionBegin;
534   elem::Conjugate(*a->emat);
535   PetscFunctionReturn(0);
536 }
537 
538 #undef __FUNCT__
539 #define __FUNCT__ "MatHermitianTranspose_Elemental"
540 static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
541 {
542   Mat            Be = *B;
543   PetscErrorCode ierr;
544   MPI_Comm       comm;
545   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
546 
547   PetscFunctionBegin;
548   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
549   /* Only out-of-place supported */
550   if (reuse == MAT_INITIAL_MATRIX){
551     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
552     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
553     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
554     ierr = MatSetUp(Be);CHKERRQ(ierr);
555     *B = Be;
556   }
557   b = (Mat_Elemental*)Be->data;
558   elem::Adjoint(*a->emat,*b->emat);
559   Be->assembled = PETSC_TRUE;
560   PetscFunctionReturn(0);
561 }
562 
563 #undef __FUNCT__
564 #define __FUNCT__ "MatSolve_Elemental"
565 static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
566 {
567   Mat_Elemental     *a = (Mat_Elemental*)A->data;
568   PetscErrorCode    ierr;
569   PetscElemScalar   *x;
570 
571   PetscFunctionBegin;
572   ierr = VecCopy(B,X);CHKERRQ(ierr);
573   ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
574   elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe;
575   xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
576   elem::DistMatrix<PetscElemScalar,elem::MC,elem::MR> xer(xe);
577   switch (A->factortype) {
578   case MAT_FACTOR_LU:
579     if ((*a->pivot).AllocatedMemory()) {
580       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,xer);
581       elem::Copy(xer,xe);
582     } else {
583       elem::lu::SolveAfter(elem::NORMAL,*a->emat,xer);
584       elem::Copy(xer,xe);
585     }
586     break;
587   case MAT_FACTOR_CHOLESKY:
588     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,xer);
589     elem::Copy(xer,xe);
590     break;
591   default:
592     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
593     break;
594   }
595   ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
596   PetscFunctionReturn(0);
597 }
598 
599 #undef __FUNCT__
600 #define __FUNCT__ "MatSolveAdd_Elemental"
601 static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
602 {
603   PetscErrorCode    ierr;
604 
605   PetscFunctionBegin;
606   ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr);
607   ierr = VecAXPY(X,1,Y);CHKERRQ(ierr);
608   PetscFunctionReturn(0);
609 }
610 
611 #undef __FUNCT__
612 #define __FUNCT__ "MatMatSolve_Elemental"
613 static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
614 {
615   Mat_Elemental *a=(Mat_Elemental*)A->data;
616   Mat_Elemental *b=(Mat_Elemental*)B->data;
617   Mat_Elemental *x=(Mat_Elemental*)X->data;
618 
619   PetscFunctionBegin;
620   elem::Copy(*b->emat,*x->emat);
621   switch (A->factortype) {
622   case MAT_FACTOR_LU:
623     if ((*a->pivot).AllocatedMemory()) {
624       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,*x->emat);
625     } else {
626       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*x->emat);
627     }
628     break;
629   case MAT_FACTOR_CHOLESKY:
630     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,*x->emat);
631     break;
632   default:
633     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
634     break;
635   }
636   PetscFunctionReturn(0);
637 }
638 
639 #undef __FUNCT__
640 #define __FUNCT__ "MatLUFactor_Elemental"
641 static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
642 {
643   Mat_Elemental  *a = (Mat_Elemental*)A->data;
644 
645   PetscFunctionBegin;
646   if (info->dtcol){
647     elem::LU(*a->emat,*a->pivot);
648   } else {
649     elem::LU(*a->emat);
650   }
651   A->factortype = MAT_FACTOR_LU;
652   A->assembled  = PETSC_TRUE;
653   PetscFunctionReturn(0);
654 }
655 
656 #undef __FUNCT__
657 #define __FUNCT__ "MatLUFactorNumeric_Elemental"
658 static PetscErrorCode  MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
659 {
660   PetscErrorCode ierr;
661 
662   PetscFunctionBegin;
663   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
664   ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr);
665   PetscFunctionReturn(0);
666 }
667 
668 #undef __FUNCT__
669 #define __FUNCT__ "MatLUFactorSymbolic_Elemental"
670 static PetscErrorCode  MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
671 {
672   PetscFunctionBegin;
673   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
674   PetscFunctionReturn(0);
675 }
676 
677 #undef __FUNCT__
678 #define __FUNCT__ "MatCholeskyFactor_Elemental"
679 static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
680 {
681   Mat_Elemental  *a = (Mat_Elemental*)A->data;
682   elem::DistMatrix<PetscElemScalar,elem::MC,elem::STAR> d;
683 
684   PetscFunctionBegin;
685   elem::Cholesky(elem::UPPER,*a->emat);
686   A->factortype = MAT_FACTOR_CHOLESKY;
687   A->assembled  = PETSC_TRUE;
688   PetscFunctionReturn(0);
689 }
690 
691 #undef __FUNCT__
692 #define __FUNCT__ "MatCholeskyFactorNumeric_Elemental"
693 static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
694 {
695   PetscErrorCode ierr;
696 
697   PetscFunctionBegin;
698   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
699   ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr);
700   PetscFunctionReturn(0);
701 }
702 
703 #undef __FUNCT__
704 #define __FUNCT__ "MatCholeskyFactorSymbolic_Elemental"
705 static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
706 {
707   PetscFunctionBegin;
708   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
709   PetscFunctionReturn(0);
710 }
711 
712 #undef __FUNCT__
713 #define __FUNCT__ "MatFactorGetSolverPackage_elemental_elemental"
714 PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type)
715 {
716   PetscFunctionBegin;
717   *type = MATSOLVERELEMENTAL;
718   PetscFunctionReturn(0);
719 }
720 
721 #undef __FUNCT__
722 #define __FUNCT__ "MatGetFactor_elemental_elemental"
723 static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
724 {
725   Mat            B;
726   PetscErrorCode ierr;
727 
728   PetscFunctionBegin;
729   /* Create the factorization matrix */
730   ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
731   ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
732   ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr);
733   ierr = MatSetUp(B);CHKERRQ(ierr);
734   B->factortype = ftype;
735   ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr);
736   *F            = B;
737   PetscFunctionReturn(0);
738 }
739 
740 #undef __FUNCT__
741 #define __FUNCT__ "MatSolverPackageRegister_Elemental"
742 PetscErrorCode MatSolverPackageRegister_Elemental(void)
743 {
744   PetscFunctionBegin;
745   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_LU,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
746   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
747   PetscFunctionReturn(0);
748 }
749 
750 #undef __FUNCT__
751 #define __FUNCT__ "MatNorm_Elemental"
752 static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
753 {
754   Mat_Elemental *a=(Mat_Elemental*)A->data;
755 
756   PetscFunctionBegin;
757   switch (type){
758   case NORM_1:
759     *nrm = elem::OneNorm(*a->emat);
760     break;
761   case NORM_FROBENIUS:
762     *nrm = elem::FrobeniusNorm(*a->emat);
763     break;
764   case NORM_INFINITY:
765     *nrm = elem::InfinityNorm(*a->emat);
766     break;
767   default:
768     printf("Error: unsupported norm type!\n");
769   }
770   PetscFunctionReturn(0);
771 }
772 
773 #undef __FUNCT__
774 #define __FUNCT__ "MatZeroEntries_Elemental"
775 static PetscErrorCode MatZeroEntries_Elemental(Mat A)
776 {
777   Mat_Elemental *a=(Mat_Elemental*)A->data;
778 
779   PetscFunctionBegin;
780   elem::Zero(*a->emat);
781   PetscFunctionReturn(0);
782 }
783 
784 #undef __FUNCT__
785 #define __FUNCT__ "MatGetOwnershipIS_Elemental"
786 static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
787 {
788   Mat_Elemental  *a = (Mat_Elemental*)A->data;
789   PetscErrorCode ierr;
790   PetscInt       i,m,shift,stride,*idx;
791 
792   PetscFunctionBegin;
793   if (rows) {
794     m = a->emat->LocalHeight();
795     shift = a->emat->ColShift();
796     stride = a->emat->ColStride();
797     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
798     for (i=0; i<m; i++) {
799       PetscInt rank,offset;
800       E2RO(A,0,shift+i*stride,&rank,&offset);
801       RO2P(A,0,rank,offset,&idx[i]);
802     }
803     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
804   }
805   if (cols) {
806     m = a->emat->LocalWidth();
807     shift = a->emat->RowShift();
808     stride = a->emat->RowStride();
809     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
810     for (i=0; i<m; i++) {
811       PetscInt rank,offset;
812       E2RO(A,1,shift+i*stride,&rank,&offset);
813       RO2P(A,1,rank,offset,&idx[i]);
814     }
815     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr);
816   }
817   PetscFunctionReturn(0);
818 }
819 
820 #undef __FUNCT__
821 #define __FUNCT__ "MatConvert_Elemental_Dense"
822 static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
823 {
824   Mat                Bmpi;
825   Mat_Elemental      *a = (Mat_Elemental*)A->data;
826   MPI_Comm           comm;
827   PetscErrorCode     ierr;
828   PetscInt           rrank,ridx,crank,cidx,nrows,ncols,i,j;
829   PetscElemScalar    v;
830   PetscBool          s1,s2,s3;
831 
832   PetscFunctionBegin;
833   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
834   ierr = PetscStrcmp(newtype,MATDENSE,&s1);CHKERRQ(ierr);
835   ierr = PetscStrcmp(newtype,MATSEQDENSE,&s2);CHKERRQ(ierr);
836   ierr = PetscStrcmp(newtype,MATMPIDENSE,&s3);CHKERRQ(ierr);
837   if (!s1 && !s2 && !s3) SETERRQ(comm,PETSC_ERR_SUP,"Unsupported New MatType: must be MATDENSE, MATSEQDENSE or MATMPIDENSE");
838   ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr);
839   ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
840   ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr);
841   ierr = MatSetUp(Bmpi);CHKERRQ(ierr);
842   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
843   for (i=0; i<nrows; i++) {
844     PetscInt erow,ecol;
845     P2RO(A,0,i,&rrank,&ridx);
846     RO2E(A,0,rrank,ridx,&erow);
847     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
848     for (j=0; j<ncols; j++) {
849       P2RO(A,1,j,&crank,&cidx);
850       RO2E(A,1,crank,cidx,&ecol);
851       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
852       v = a->emat->Get(erow,ecol);
853       ierr = MatSetValues(Bmpi,1,&i,1,&j,(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
854     }
855   }
856   ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
857   ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
858   if (reuse == MAT_REUSE_MATRIX) {
859     ierr = MatHeaderReplace(A,Bmpi);CHKERRQ(ierr);
860   } else {
861     *B = Bmpi;
862   }
863   PetscFunctionReturn(0);
864 }
865 
866 #undef __FUNCT__
867 #define __FUNCT__ "MatConvert_SeqAIJ_Elemental"
868 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
869 {
870   Mat               mat_elemental;
871   PetscErrorCode    ierr;
872   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols;
873   const PetscInt    *cols;
874   const PetscScalar *vals;
875 
876   PetscFunctionBegin;
877   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
878   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
879   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
880   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
881   for (row=0; row<M; row++) {
882     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
883     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
884     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
885     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
886   }
887   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
888   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
889 
890   if (reuse == MAT_REUSE_MATRIX) {
891     ierr = MatHeaderReplace(A,mat_elemental);CHKERRQ(ierr);
892   } else {
893     *newmat = mat_elemental;
894   }
895   PetscFunctionReturn(0);
896 }
897 
898 #undef __FUNCT__
899 #define __FUNCT__ "MatConvert_MPIAIJ_Elemental"
900 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
901 {
902   Mat               mat_elemental;
903   PetscErrorCode    ierr;
904   PetscInt          row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j;
905   const PetscInt    *cols;
906   const PetscScalar *vals;
907 
908   PetscFunctionBegin;
909   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
910   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);CHKERRQ(ierr);
911   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
912   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
913   for (row=rstart; row<rend; row++) {
914     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
915     for (j=0; j<ncols; j++) {
916       /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
917       ierr = MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);CHKERRQ(ierr);
918     }
919     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
920   }
921   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
922   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
923 
924   if (reuse == MAT_REUSE_MATRIX) {
925     ierr = MatHeaderReplace(A,mat_elemental);CHKERRQ(ierr);
926   } else {
927     *newmat = mat_elemental;
928   }
929   PetscFunctionReturn(0);
930 }
931 
932 #undef __FUNCT__
933 #define __FUNCT__ "MatDestroy_Elemental"
934 static PetscErrorCode MatDestroy_Elemental(Mat A)
935 {
936   Mat_Elemental      *a = (Mat_Elemental*)A->data;
937   PetscErrorCode     ierr;
938   Mat_Elemental_Grid *commgrid;
939   PetscBool          flg;
940   MPI_Comm           icomm;
941 
942   PetscFunctionBegin;
943   a->interface->Detach();
944   delete a->interface;
945   delete a->esubmat;
946   delete a->emat;
947 
948   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
949   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
950   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
951   if (--commgrid->grid_refct == 0) {
952     delete commgrid->grid;
953     ierr = PetscFree(commgrid);CHKERRQ(ierr);
954   }
955   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
956   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr);
957   ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr);
958   ierr = PetscFree(A->data);CHKERRQ(ierr);
959   PetscFunctionReturn(0);
960 }
961 
962 #undef __FUNCT__
963 #define __FUNCT__ "MatSetUp_Elemental"
964 PetscErrorCode MatSetUp_Elemental(Mat A)
965 {
966   Mat_Elemental  *a = (Mat_Elemental*)A->data;
967   PetscErrorCode ierr;
968   PetscMPIInt    rsize,csize;
969 
970   PetscFunctionBegin;
971   ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
972   ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
973 
974   a->emat->Resize(A->rmap->N,A->cmap->N);CHKERRQ(ierr);
975   elem::Zero(*a->emat);
976 
977   ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr);
978   ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr);
979   if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
980   a->commsize = rsize;
981   a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
982   a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
983   a->m[0]  = A->rmap->N / rsize + (a->mr[0] != rsize);
984   a->m[1]  = A->cmap->N / csize + (a->mr[1] != csize);
985   PetscFunctionReturn(0);
986 }
987 
988 #undef __FUNCT__
989 #define __FUNCT__ "MatAssemblyBegin_Elemental"
990 PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
991 {
992   Mat_Elemental  *a = (Mat_Elemental*)A->data;
993 
994   PetscFunctionBegin;
995   a->interface->Detach();
996   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
997   PetscFunctionReturn(0);
998 }
999 
1000 #undef __FUNCT__
1001 #define __FUNCT__ "MatAssemblyEnd_Elemental"
1002 PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
1003 {
1004   PetscFunctionBegin;
1005   /* Currently does nothing */
1006   PetscFunctionReturn(0);
1007 }
1008 
1009 /* -------------------------------------------------------------------*/
1010 static struct _MatOps MatOps_Values = {
1011        MatSetValues_Elemental,
1012        0,
1013        0,
1014        MatMult_Elemental,
1015 /* 4*/ MatMultAdd_Elemental,
1016        MatMultTranspose_Elemental,
1017        MatMultTransposeAdd_Elemental,
1018        MatSolve_Elemental,
1019        MatSolveAdd_Elemental,
1020        0, //MatSolveTranspose_Elemental,
1021 /*10*/ 0, //MatSolveTransposeAdd_Elemental,
1022        MatLUFactor_Elemental,
1023        MatCholeskyFactor_Elemental,
1024        0,
1025        MatTranspose_Elemental,
1026 /*15*/ MatGetInfo_Elemental,
1027        0,
1028        MatGetDiagonal_Elemental,
1029        MatDiagonalScale_Elemental,
1030        MatNorm_Elemental,
1031 /*20*/ MatAssemblyBegin_Elemental,
1032        MatAssemblyEnd_Elemental,
1033        0, //MatSetOption_Elemental,
1034        MatZeroEntries_Elemental,
1035 /*24*/ 0,
1036        MatLUFactorSymbolic_Elemental,
1037        MatLUFactorNumeric_Elemental,
1038        MatCholeskyFactorSymbolic_Elemental,
1039        MatCholeskyFactorNumeric_Elemental,
1040 /*29*/ MatSetUp_Elemental,
1041        0,
1042        0,
1043        0,
1044        0,
1045 /*34*/ MatDuplicate_Elemental,
1046        0,
1047        0,
1048        0,
1049        0,
1050 /*39*/ MatAXPY_Elemental,
1051        0,
1052        0,
1053        0,
1054        MatCopy_Elemental,
1055 /*44*/ 0,
1056        MatScale_Elemental,
1057        0,
1058        0,
1059        0,
1060 /*49*/ 0,
1061        0,
1062        0,
1063        0,
1064        0,
1065 /*54*/ 0,
1066        0,
1067        0,
1068        0,
1069        0,
1070 /*59*/ 0,
1071        MatDestroy_Elemental,
1072        MatView_Elemental,
1073        0,
1074        0,
1075 /*64*/ 0,
1076        0,
1077        0,
1078        0,
1079        0,
1080 /*69*/ 0,
1081        0,
1082        MatConvert_Elemental_Dense,
1083        0,
1084        0,
1085 /*74*/ 0,
1086        0,
1087        0,
1088        0,
1089        0,
1090 /*79*/ 0,
1091        0,
1092        0,
1093        0,
1094        0,
1095 /*84*/ 0,
1096        0,
1097        0,
1098        0,
1099        0,
1100 /*89*/ MatMatMult_Elemental,
1101        MatMatMultSymbolic_Elemental,
1102        MatMatMultNumeric_Elemental,
1103        0,
1104        0,
1105 /*94*/ 0,
1106        MatMatTransposeMult_Elemental,
1107        MatMatTransposeMultSymbolic_Elemental,
1108        MatMatTransposeMultNumeric_Elemental,
1109        0,
1110 /*99*/ 0,
1111        0,
1112        0,
1113        MatConjugate_Elemental,
1114        0,
1115 /*104*/0,
1116        0,
1117        0,
1118        0,
1119        0,
1120 /*109*/MatMatSolve_Elemental,
1121        0,
1122        0,
1123        0,
1124        0,
1125 /*114*/0,
1126        0,
1127        0,
1128        0,
1129        0,
1130 /*119*/0,
1131        MatHermitianTranspose_Elemental,
1132        0,
1133        0,
1134        0,
1135 /*124*/0,
1136        0,
1137        0,
1138        0,
1139        0,
1140 /*129*/0,
1141        0,
1142        0,
1143        0,
1144        0,
1145 /*134*/0,
1146        0,
1147        0,
1148        0,
1149        0
1150 };
1151 
1152 /*MC
1153    MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1154 
1155    Options Database Keys:
1156 + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
1157 - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1158 
1159   Level: beginner
1160 
1161 .seealso: MATDENSE
1162 M*/
1163 
1164 #undef __FUNCT__
1165 #define __FUNCT__ "MatCreate_Elemental"
1166 PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1167 {
1168   Mat_Elemental      *a;
1169   PetscErrorCode     ierr;
1170   PetscBool          flg,flg1;
1171   Mat_Elemental_Grid *commgrid;
1172   MPI_Comm           icomm;
1173   PetscInt           optv1;
1174 
1175   PetscFunctionBegin;
1176   ierr = PetscElementalInitializePackage();CHKERRQ(ierr);
1177   ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
1178   A->insertmode = NOT_SET_VALUES;
1179 
1180   ierr = PetscNewLog(A,&a);CHKERRQ(ierr);
1181   A->data = (void*)a;
1182 
1183   /* Set up the elemental matrix */
1184   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
1185 
1186   /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
1187   if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
1188     ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);
1189   }
1190   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
1191   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
1192   if (!flg) {
1193     ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr);
1194 
1195     ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr);
1196     /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until elem::Grid() is called */
1197     ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",elem::mpi::Size(cxxcomm),&optv1,&flg1);CHKERRQ(ierr);
1198     if (flg1) {
1199       if (elem::mpi::Size(cxxcomm) % optv1 != 0) {
1200         SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)elem::mpi::Size(cxxcomm));
1201       }
1202       commgrid->grid = new elem::Grid(cxxcomm,optv1); /* use user-provided grid height */
1203     } else {
1204       commgrid->grid = new elem::Grid(cxxcomm); /* use Elemental default grid sizes */
1205     }
1206     commgrid->grid_refct = 1;
1207     ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr);
1208     PetscOptionsEnd();
1209   } else {
1210     commgrid->grid_refct++;
1211   }
1212   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
1213   a->grid      = commgrid->grid;
1214   a->emat      = new elem::DistMatrix<PetscElemScalar>(*a->grid);
1215   a->esubmat   = new elem::Matrix<PetscElemScalar>(1,1);
1216   a->interface = new elem::AxpyInterface<PetscElemScalar>;
1217   a->pivot     = new elem::DistMatrix<PetscInt,elem::VC,elem::STAR>;
1218 
1219   /* build cache for off array entries formed */
1220   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
1221 
1222   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr);
1223 
1224   ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr);
1225   PetscFunctionReturn(0);
1226 }
1227 
1228