xref: /petsc/src/mat/impls/elemental/matelem.cxx (revision cdc753b64c658d532d7813e942002e67025667a6)
1db31f6deSJed Brown #include <../src/mat/impls/elemental/matelemimpl.h> /*I "petscmat.h" I*/
2db31f6deSJed Brown 
35e9f5b67SHong Zhang /*
45e9f5b67SHong Zhang     The variable Petsc_Elemental_keyval is used to indicate an MPI attribute that
55e9f5b67SHong Zhang   is attached to a communicator, in this case the attribute is a Mat_Elemental_Grid
65e9f5b67SHong Zhang */
75e9f5b67SHong Zhang static PetscMPIInt Petsc_Elemental_keyval = MPI_KEYVAL_INVALID;
85e9f5b67SHong Zhang 
9db31f6deSJed Brown /*@C
10db31f6deSJed Brown    PetscElementalInitializePackage - Initialize Elemental package
11db31f6deSJed Brown 
12db31f6deSJed Brown    Logically Collective
13db31f6deSJed Brown 
14db31f6deSJed Brown    Level: developer
15db31f6deSJed Brown 
16db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalFinalizePackage()
17db31f6deSJed Brown @*/
18607a6623SBarry Smith PetscErrorCode PetscElementalInitializePackage(void)
19db31f6deSJed Brown {
20db31f6deSJed Brown   PetscErrorCode ierr;
21db31f6deSJed Brown 
22db31f6deSJed Brown   PetscFunctionBegin;
23e8146892SSatish Balay   if (El::Initialized()) PetscFunctionReturn(0);
24e05a9d59SSatish Balay   El::Initialize();   /* called by the 1st call of MatCreate_Elemental */
25db31f6deSJed Brown   ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr);
26db31f6deSJed Brown   PetscFunctionReturn(0);
27db31f6deSJed Brown }
28db31f6deSJed Brown 
29db31f6deSJed Brown /*@C
30db31f6deSJed Brown    PetscElementalFinalizePackage - Finalize Elemental package
31db31f6deSJed Brown 
32db31f6deSJed Brown    Logically Collective
33db31f6deSJed Brown 
34db31f6deSJed Brown    Level: developer
35db31f6deSJed Brown 
36db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage()
37db31f6deSJed Brown @*/
38db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void)
39db31f6deSJed Brown {
40db31f6deSJed Brown   PetscFunctionBegin;
41e8146892SSatish Balay   El::Finalize();  /* called by PetscFinalize() */
42db31f6deSJed Brown   PetscFunctionReturn(0);
43db31f6deSJed Brown }
44db31f6deSJed Brown 
45db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
46db31f6deSJed Brown {
47db31f6deSJed Brown   PetscErrorCode ierr;
48db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
49db31f6deSJed Brown   PetscBool      iascii;
50db31f6deSJed Brown 
51db31f6deSJed Brown   PetscFunctionBegin;
52db31f6deSJed Brown   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
53db31f6deSJed Brown   if (iascii) {
54db31f6deSJed Brown     PetscViewerFormat format;
55db31f6deSJed Brown     ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
56db31f6deSJed Brown     if (format == PETSC_VIEWER_ASCII_INFO) {
5779673f7bSHong Zhang       /* call elemental viewing function */
582d8adcc7SHong Zhang       ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr);
59ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr);
60ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr);
614fe7bbcaSHong Zhang       if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
6279673f7bSHong Zhang         /* call elemental viewing function */
63ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr);
644fe7bbcaSHong Zhang       }
6579673f7bSHong Zhang 
66db31f6deSJed Brown     } else if (format == PETSC_VIEWER_DEFAULT) {
67db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr);
68e8146892SSatish Balay       El::Print( *a->emat, "Elemental matrix (cyclic ordering)" );
69db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr);
70834d3fecSHong Zhang       if (A->factortype == MAT_FACTOR_NONE){
7161119200SXuan Zhou         Mat Adense;
72ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
7361119200SXuan Zhou         ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
7461119200SXuan Zhou         ierr = MatView(Adense,viewer);CHKERRQ(ierr);
7561119200SXuan Zhou         ierr = MatDestroy(&Adense);CHKERRQ(ierr);
76834d3fecSHong Zhang       }
77ce94432eSBarry Smith     } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
78d2daa67eSHong Zhang   } else {
795cb544a0SHong Zhang     /* convert to dense format and call MatView() */
8061119200SXuan Zhou     Mat Adense;
81ce94432eSBarry Smith     ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
8261119200SXuan Zhou     ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
8361119200SXuan Zhou     ierr = MatView(Adense,viewer);CHKERRQ(ierr);
8461119200SXuan Zhou     ierr = MatDestroy(&Adense);CHKERRQ(ierr);
85d2daa67eSHong Zhang   }
86db31f6deSJed Brown   PetscFunctionReturn(0);
87db31f6deSJed Brown }
88db31f6deSJed Brown 
8915767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
90180a43e4SHong Zhang {
9115767789SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
9215767789SHong Zhang 
93180a43e4SHong Zhang   PetscFunctionBegin;
945cb544a0SHong Zhang   info->block_size = 1.0;
9515767789SHong Zhang 
9615767789SHong Zhang   if (flag == MAT_LOCAL) {
9715767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
9815767789SHong Zhang     info->nz_used        = info->nz_allocated;
9915767789SHong Zhang   } else if (flag == MAT_GLOBAL_MAX) {
100b2566f29SBarry Smith     //ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr);
10115767789SHong Zhang     /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
10215767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
10315767789SHong Zhang   } else if (flag == MAT_GLOBAL_SUM) {
10415767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
10515767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
10615767789SHong Zhang     info->nz_used        = info->nz_allocated; /* assume Elemental does accurate allocation */
107b2566f29SBarry Smith     //ierr = MPIU_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr);
10815767789SHong Zhang     //PetscPrintf(PETSC_COMM_SELF,"    ... [%d] locally allocated %g\n",rank,info->nz_allocated);
10915767789SHong Zhang   }
11015767789SHong Zhang 
11115767789SHong Zhang   info->nz_unneeded       = 0.0;
11215767789SHong Zhang   info->assemblies        = (double)A->num_ass;
11315767789SHong Zhang   info->mallocs           = 0;
11415767789SHong Zhang   info->memory            = ((PetscObject)A)->mem;
11515767789SHong Zhang   info->fill_ratio_given  = 0; /* determined by Elemental */
11615767789SHong Zhang   info->fill_ratio_needed = 0;
11715767789SHong Zhang   info->factor_mallocs    = 0;
118180a43e4SHong Zhang   PetscFunctionReturn(0);
119180a43e4SHong Zhang }
120180a43e4SHong Zhang 
12132d7a744SHong Zhang PetscErrorCode MatSetOption_Elemental(Mat A,MatOption op,PetscBool flg)
12232d7a744SHong Zhang {
12332d7a744SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
12432d7a744SHong Zhang 
12532d7a744SHong Zhang   PetscFunctionBegin;
12632d7a744SHong Zhang   switch (op) {
12732d7a744SHong Zhang   case MAT_NEW_NONZERO_LOCATIONS:
12832d7a744SHong Zhang   case MAT_NEW_NONZERO_LOCATION_ERR:
12932d7a744SHong Zhang   case MAT_NEW_NONZERO_ALLOCATION_ERR:
130891a0571SHong Zhang   case MAT_SYMMETRIC:
131891a0571SHong Zhang     break;
13232d7a744SHong Zhang   case MAT_ROW_ORIENTED:
13332d7a744SHong Zhang     a->roworiented = flg;
13432d7a744SHong Zhang     break;
13532d7a744SHong Zhang   default:
13632d7a744SHong Zhang     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %s",MatOptions[op]);
13732d7a744SHong Zhang   }
13832d7a744SHong Zhang   PetscFunctionReturn(0);
13932d7a744SHong Zhang }
14032d7a744SHong Zhang 
141e6dea9dbSXuan Zhou static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode)
142db31f6deSJed Brown {
143db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
144fbbcb730SJack Poulson   PetscInt       i,j,rrank,ridx,crank,cidx,erow,ecol,numQueues=0;
145db31f6deSJed Brown 
146db31f6deSJed Brown   PetscFunctionBegin;
147fbbcb730SJack Poulson   // TODO: Initialize matrix to all zeros?
148fbbcb730SJack Poulson 
149fbbcb730SJack Poulson   // Count the number of queues from this process
15032d7a744SHong Zhang   if (a->roworiented) {
151db31f6deSJed Brown     for (i=0; i<nr; i++) {
152db31f6deSJed Brown       if (rows[i] < 0) continue;
153db31f6deSJed Brown       P2RO(A,0,rows[i],&rrank,&ridx);
154db31f6deSJed Brown       RO2E(A,0,rrank,ridx,&erow);
155ce94432eSBarry Smith       if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
156db31f6deSJed Brown       for (j=0; j<nc; j++) {
157db31f6deSJed Brown         if (cols[j] < 0) continue;
158db31f6deSJed Brown         P2RO(A,1,cols[j],&crank,&cidx);
159db31f6deSJed Brown         RO2E(A,1,crank,cidx,&ecol);
160ce94432eSBarry Smith         if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
161fbbcb730SJack Poulson         if (!a->emat->IsLocal(erow,ecol) ){ /* off-proc entry */
162fbbcb730SJack Poulson           /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */
163aae2c449SHong Zhang           if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
164fbbcb730SJack Poulson           ++numQueues;
165aae2c449SHong Zhang           continue;
166ed36708cSHong Zhang         }
167fbbcb730SJack Poulson         /* printf("Locally updating (%d,%d)\n",erow,ecol); */
168db31f6deSJed Brown         switch (imode) {
169fbbcb730SJack Poulson         case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break;
170fbbcb730SJack Poulson         case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break;
171ce94432eSBarry Smith         default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
172db31f6deSJed Brown         }
173db31f6deSJed Brown       }
174db31f6deSJed Brown     }
175fbbcb730SJack Poulson 
176fbbcb730SJack Poulson     /* printf("numQueues=%d\n",numQueues); */
177fbbcb730SJack Poulson     a->emat->Reserve( numQueues );
178fbbcb730SJack Poulson     for (i=0; i<nr; i++) {
179fbbcb730SJack Poulson       if (rows[i] < 0) continue;
180fbbcb730SJack Poulson       P2RO(A,0,rows[i],&rrank,&ridx);
181fbbcb730SJack Poulson       RO2E(A,0,rrank,ridx,&erow);
182fbbcb730SJack Poulson       for (j=0; j<nc; j++) {
183fbbcb730SJack Poulson         if (cols[j] < 0) continue;
184fbbcb730SJack Poulson         P2RO(A,1,cols[j],&crank,&cidx);
185fbbcb730SJack Poulson         RO2E(A,1,crank,cidx,&ecol);
186fbbcb730SJack Poulson         if ( !a->emat->IsLocal(erow,ecol) ) { /*off-proc entry*/
187fbbcb730SJack Poulson           /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */
188fbbcb730SJack Poulson           a->emat->QueueUpdate( erow, ecol, vals[i*nc+j] );
189fbbcb730SJack Poulson         }
190fbbcb730SJack Poulson       }
191fbbcb730SJack Poulson     }
19232d7a744SHong Zhang   } else { /* columnoriented */
19332d7a744SHong Zhang     for (j=0; j<nc; j++) {
19432d7a744SHong Zhang       if (cols[j] < 0) continue;
19532d7a744SHong Zhang       P2RO(A,1,cols[j],&crank,&cidx);
19632d7a744SHong Zhang       RO2E(A,1,crank,cidx,&ecol);
19732d7a744SHong Zhang       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
19832d7a744SHong Zhang       for (i=0; i<nr; i++) {
19932d7a744SHong Zhang         if (rows[i] < 0) continue;
20032d7a744SHong Zhang         P2RO(A,0,rows[i],&rrank,&ridx);
20132d7a744SHong Zhang         RO2E(A,0,rrank,ridx,&erow);
20232d7a744SHong Zhang         if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
20332d7a744SHong Zhang         if (!a->emat->IsLocal(erow,ecol) ){ /* off-proc entry */
20432d7a744SHong Zhang           /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */
20532d7a744SHong Zhang           if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
20632d7a744SHong Zhang           ++numQueues;
20732d7a744SHong Zhang           continue;
20832d7a744SHong Zhang         }
20932d7a744SHong Zhang         /* printf("Locally updating (%d,%d)\n",erow,ecol); */
21032d7a744SHong Zhang         switch (imode) {
21132d7a744SHong Zhang         case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break;
21232d7a744SHong Zhang         case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break;
21332d7a744SHong Zhang         default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
21432d7a744SHong Zhang         }
21532d7a744SHong Zhang       }
21632d7a744SHong Zhang     }
21732d7a744SHong Zhang 
21832d7a744SHong Zhang     /* printf("numQueues=%d\n",numQueues); */
21932d7a744SHong Zhang     a->emat->Reserve( numQueues );
22032d7a744SHong Zhang     for (j=0; j<nc; j++) {
22132d7a744SHong Zhang       if (cols[j] < 0) continue;
22232d7a744SHong Zhang       P2RO(A,1,cols[j],&crank,&cidx);
22332d7a744SHong Zhang       RO2E(A,1,crank,cidx,&ecol);
22432d7a744SHong Zhang 
22532d7a744SHong Zhang       for (i=0; i<nr; i++) {
22632d7a744SHong Zhang         if (rows[i] < 0) continue;
22732d7a744SHong Zhang         P2RO(A,0,rows[i],&rrank,&ridx);
22832d7a744SHong Zhang         RO2E(A,0,rrank,ridx,&erow);
22932d7a744SHong Zhang         if ( !a->emat->IsLocal(erow,ecol) ) { /*off-proc entry*/
23032d7a744SHong Zhang           /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */
23132d7a744SHong Zhang           a->emat->QueueUpdate( erow, ecol, vals[i+j*nr] );
23232d7a744SHong Zhang         }
23332d7a744SHong Zhang       }
23432d7a744SHong Zhang     }
23532d7a744SHong Zhang   }
236db31f6deSJed Brown   PetscFunctionReturn(0);
237db31f6deSJed Brown }
238db31f6deSJed Brown 
239db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
240db31f6deSJed Brown {
241db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
242db31f6deSJed Brown   PetscErrorCode        ierr;
243e6dea9dbSXuan Zhou   const PetscElemScalar *x;
244e6dea9dbSXuan Zhou   PetscElemScalar       *y;
245df311e6cSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
246db31f6deSJed Brown 
247db31f6deSJed Brown   PetscFunctionBegin;
248e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
249e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
250db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
251e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
2520c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
2530c18141cSBarry Smith     ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n);
254e8146892SSatish Balay     El::Gemv(El::NORMAL,one,*a->emat,xe,zero,ye);
255db31f6deSJed Brown   }
256e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
257e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
258db31f6deSJed Brown   PetscFunctionReturn(0);
259db31f6deSJed Brown }
260db31f6deSJed Brown 
2619426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
2629426833fSXuan Zhou {
2639426833fSXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
2649426833fSXuan Zhou   PetscErrorCode        ierr;
265df311e6cSXuan Zhou   const PetscElemScalar *x;
266df311e6cSXuan Zhou   PetscElemScalar       *y;
267e6dea9dbSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
2689426833fSXuan Zhou 
2699426833fSXuan Zhou   PetscFunctionBegin;
270e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
271e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2729426833fSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
273e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
2740c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
2750c18141cSBarry Smith     ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n);
276e8146892SSatish Balay     El::Gemv(El::TRANSPOSE,one,*a->emat,xe,zero,ye);
2779426833fSXuan Zhou   }
278e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
279e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2809426833fSXuan Zhou   PetscFunctionReturn(0);
2819426833fSXuan Zhou }
2829426833fSXuan Zhou 
283db31f6deSJed Brown static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
284db31f6deSJed Brown {
285db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
286db31f6deSJed Brown   PetscErrorCode        ierr;
287df311e6cSXuan Zhou   const PetscElemScalar *x;
288df311e6cSXuan Zhou   PetscElemScalar       *z;
289e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
290db31f6deSJed Brown 
291db31f6deSJed Brown   PetscFunctionBegin;
292db31f6deSJed Brown   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
293e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
294e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
295db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
296e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
2970c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
2980c18141cSBarry Smith     ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n);
299e8146892SSatish Balay     El::Gemv(El::NORMAL,one,*a->emat,xe,one,ze);
300db31f6deSJed Brown   }
301e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
302e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
303db31f6deSJed Brown   PetscFunctionReturn(0);
304db31f6deSJed Brown }
305db31f6deSJed Brown 
306e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
307e883f9d5SXuan Zhou {
308e883f9d5SXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
309e883f9d5SXuan Zhou   PetscErrorCode        ierr;
310df311e6cSXuan Zhou   const PetscElemScalar *x;
311df311e6cSXuan Zhou   PetscElemScalar       *z;
312e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
313e883f9d5SXuan Zhou 
314e883f9d5SXuan Zhou   PetscFunctionBegin;
315e883f9d5SXuan Zhou   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
316e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
317e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
318e883f9d5SXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
319e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
3200c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
3210c18141cSBarry Smith     ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n);
322e8146892SSatish Balay     El::Gemv(El::TRANSPOSE,one,*a->emat,xe,one,ze);
323e883f9d5SXuan Zhou   }
324e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
325e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
326e883f9d5SXuan Zhou   PetscFunctionReturn(0);
327e883f9d5SXuan Zhou }
328e883f9d5SXuan Zhou 
3299a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
330c1d1b975SXuan Zhou {
331c1d1b975SXuan Zhou   Mat_Elemental    *a = (Mat_Elemental*)A->data;
332c1d1b975SXuan Zhou   Mat_Elemental    *b = (Mat_Elemental*)B->data;
3339a9e8502SHong Zhang   Mat_Elemental    *c = (Mat_Elemental*)C->data;
334e6dea9dbSXuan Zhou   PetscElemScalar  one = 1,zero = 0;
335c1d1b975SXuan Zhou 
336c1d1b975SXuan Zhou   PetscFunctionBegin;
337aae2c449SHong Zhang   { /* Scoping so that constructor is called before pointer is returned */
338e8146892SSatish Balay     El::Gemm(El::NORMAL,El::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
339aae2c449SHong Zhang   }
3409a9e8502SHong Zhang   C->assembled = PETSC_TRUE;
3419a9e8502SHong Zhang   PetscFunctionReturn(0);
3429a9e8502SHong Zhang }
3439a9e8502SHong Zhang 
3449a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
3459a9e8502SHong Zhang {
3469a9e8502SHong Zhang   PetscErrorCode ierr;
3479a9e8502SHong Zhang   Mat            Ce;
348ce94432eSBarry Smith   MPI_Comm       comm;
3499a9e8502SHong Zhang 
3509a9e8502SHong Zhang   PetscFunctionBegin;
351ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
3529a9e8502SHong Zhang   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
3539a9e8502SHong Zhang   ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
3549a9e8502SHong Zhang   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
3559a9e8502SHong Zhang   ierr = MatSetUp(Ce);CHKERRQ(ierr);
3569a9e8502SHong Zhang   *C = Ce;
3579a9e8502SHong Zhang   PetscFunctionReturn(0);
3589a9e8502SHong Zhang }
3599a9e8502SHong Zhang 
3609a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3619a9e8502SHong Zhang {
3629a9e8502SHong Zhang   PetscErrorCode ierr;
3639a9e8502SHong Zhang 
3649a9e8502SHong Zhang   PetscFunctionBegin;
3659a9e8502SHong Zhang   if (scall == MAT_INITIAL_MATRIX){
3663ff4c91cSHong Zhang     ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3679a9e8502SHong Zhang     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
3683ff4c91cSHong Zhang     ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3699a9e8502SHong Zhang   }
3703ff4c91cSHong Zhang   ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
3719a9e8502SHong Zhang   ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
3723ff4c91cSHong Zhang   ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
373c1d1b975SXuan Zhou   PetscFunctionReturn(0);
374c1d1b975SXuan Zhou }
375c1d1b975SXuan Zhou 
376df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
377df311e6cSXuan Zhou {
378df311e6cSXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
379df311e6cSXuan Zhou   Mat_Elemental      *b = (Mat_Elemental*)B->data;
380df311e6cSXuan Zhou   Mat_Elemental      *c = (Mat_Elemental*)C->data;
381e6dea9dbSXuan Zhou   PetscElemScalar    one = 1,zero = 0;
382df311e6cSXuan Zhou 
383df311e6cSXuan Zhou   PetscFunctionBegin;
384df311e6cSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
385e8146892SSatish Balay     El::Gemm(El::NORMAL,El::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
386df311e6cSXuan Zhou   }
387df311e6cSXuan Zhou   C->assembled = PETSC_TRUE;
388df311e6cSXuan Zhou   PetscFunctionReturn(0);
389df311e6cSXuan Zhou }
390df311e6cSXuan Zhou 
391df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
392df311e6cSXuan Zhou {
393df311e6cSXuan Zhou   PetscErrorCode ierr;
394df311e6cSXuan Zhou   Mat            Ce;
395ce94432eSBarry Smith   MPI_Comm       comm;
396df311e6cSXuan Zhou 
397df311e6cSXuan Zhou   PetscFunctionBegin;
398ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
399df311e6cSXuan Zhou   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
400df311e6cSXuan Zhou   ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
401df311e6cSXuan Zhou   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
402df311e6cSXuan Zhou   ierr = MatSetUp(Ce);CHKERRQ(ierr);
403df311e6cSXuan Zhou   *C = Ce;
404df311e6cSXuan Zhou   PetscFunctionReturn(0);
405df311e6cSXuan Zhou }
406df311e6cSXuan Zhou 
407df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
408df311e6cSXuan Zhou {
409df311e6cSXuan Zhou   PetscErrorCode ierr;
410df311e6cSXuan Zhou 
411df311e6cSXuan Zhou   PetscFunctionBegin;
412df311e6cSXuan Zhou   if (scall == MAT_INITIAL_MATRIX){
413df311e6cSXuan Zhou     ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
414df311e6cSXuan Zhou     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
415df311e6cSXuan Zhou     ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
416df311e6cSXuan Zhou   }
417df311e6cSXuan Zhou   ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
418df311e6cSXuan Zhou   ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
419df311e6cSXuan Zhou   ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
420df311e6cSXuan Zhou   PetscFunctionReturn(0);
421df311e6cSXuan Zhou }
422df311e6cSXuan Zhou 
423a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
42461119200SXuan Zhou {
425a9d89745SXuan Zhou   PetscInt        i,nrows,ncols,nD,rrank,ridx,crank,cidx;
426a9d89745SXuan Zhou   Mat_Elemental   *a = (Mat_Elemental*)A->data;
42761119200SXuan Zhou   PetscErrorCode  ierr;
428a9d89745SXuan Zhou   PetscElemScalar v;
429ce94432eSBarry Smith   MPI_Comm        comm;
43061119200SXuan Zhou 
43161119200SXuan Zhou   PetscFunctionBegin;
432ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
433a9d89745SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
434a9d89745SXuan Zhou   nD = nrows>ncols ? ncols : nrows;
435a9d89745SXuan Zhou   for (i=0; i<nD; i++) {
436a9d89745SXuan Zhou     PetscInt erow,ecol;
437a9d89745SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
438a9d89745SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
439a9d89745SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
440a9d89745SXuan Zhou     P2RO(A,1,i,&crank,&cidx);
441a9d89745SXuan Zhou     RO2E(A,1,crank,cidx,&ecol);
442a9d89745SXuan Zhou     if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
443a9d89745SXuan Zhou     v = a->emat->Get(erow,ecol);
4447c8b904fSSatish Balay     ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr);
445ade3cc5eSXuan Zhou   }
446a9d89745SXuan Zhou   ierr = VecAssemblyBegin(D);CHKERRQ(ierr);
447a9d89745SXuan Zhou   ierr = VecAssemblyEnd(D);CHKERRQ(ierr);
44861119200SXuan Zhou   PetscFunctionReturn(0);
44961119200SXuan Zhou }
45061119200SXuan Zhou 
451ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
452ade3cc5eSXuan Zhou {
453ade3cc5eSXuan Zhou   Mat_Elemental         *x = (Mat_Elemental*)X->data;
454ade3cc5eSXuan Zhou   const PetscElemScalar *d;
455ade3cc5eSXuan Zhou   PetscErrorCode        ierr;
456ade3cc5eSXuan Zhou 
457ade3cc5eSXuan Zhou   PetscFunctionBegin;
4589065cd98SJed Brown   if (R) {
459ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
460e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
4610c18141cSBarry Smith     de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n);
462e8146892SSatish Balay     El::DiagonalScale(El::RIGHT,El::NORMAL,de,*x->emat);
463ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
4649065cd98SJed Brown   }
4659065cd98SJed Brown   if (L) {
466ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
467e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
4680c18141cSBarry Smith     de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n);
469e8146892SSatish Balay     El::DiagonalScale(El::LEFT,El::NORMAL,de,*x->emat);
470ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
471ade3cc5eSXuan Zhou   }
472ade3cc5eSXuan Zhou   PetscFunctionReturn(0);
473ade3cc5eSXuan Zhou }
474ade3cc5eSXuan Zhou 
475e6dea9dbSXuan Zhou static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a)
47665b78793SXuan Zhou {
47765b78793SXuan Zhou   Mat_Elemental  *x = (Mat_Elemental*)X->data;
47865b78793SXuan Zhou 
47965b78793SXuan Zhou   PetscFunctionBegin;
480e8146892SSatish Balay   El::Scale((PetscElemScalar)a,*x->emat);
48165b78793SXuan Zhou   PetscFunctionReturn(0);
48265b78793SXuan Zhou }
48365b78793SXuan Zhou 
484f82baa17SHong Zhang /*
485f82baa17SHong Zhang   MatAXPY - Computes Y = a*X + Y.
486f82baa17SHong Zhang */
487e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
488e09a3074SHong Zhang {
489e09a3074SHong Zhang   Mat_Elemental  *x = (Mat_Elemental*)X->data;
490e09a3074SHong Zhang   Mat_Elemental  *y = (Mat_Elemental*)Y->data;
49168446cf8SJed Brown   PetscErrorCode ierr;
492e09a3074SHong Zhang 
493e09a3074SHong Zhang   PetscFunctionBegin;
494e8146892SSatish Balay   El::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
49521f6c9c4SJed Brown   ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr);
496e09a3074SHong Zhang   PetscFunctionReturn(0);
497e09a3074SHong Zhang }
498e09a3074SHong Zhang 
499d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
500d6223691SXuan Zhou {
501d6223691SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
502d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
503d6223691SXuan Zhou 
504d6223691SXuan Zhou   PetscFunctionBegin;
505e8146892SSatish Balay   El::Copy(*a->emat,*b->emat);
506*cdc753b6SBarry Smith   ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr);
507d6223691SXuan Zhou   PetscFunctionReturn(0);
508d6223691SXuan Zhou }
509d6223691SXuan Zhou 
510df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
511df311e6cSXuan Zhou {
512df311e6cSXuan Zhou   Mat            Be;
513ce94432eSBarry Smith   MPI_Comm       comm;
514df311e6cSXuan Zhou   Mat_Elemental  *a=(Mat_Elemental*)A->data;
515df311e6cSXuan Zhou   PetscErrorCode ierr;
516df311e6cSXuan Zhou 
517df311e6cSXuan Zhou   PetscFunctionBegin;
518ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
519df311e6cSXuan Zhou   ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
520df311e6cSXuan Zhou   ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
521df311e6cSXuan Zhou   ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
522df311e6cSXuan Zhou   ierr = MatSetUp(Be);CHKERRQ(ierr);
523df311e6cSXuan Zhou   *B = Be;
524df311e6cSXuan Zhou   if (op == MAT_COPY_VALUES) {
525df311e6cSXuan Zhou     Mat_Elemental *b=(Mat_Elemental*)Be->data;
526e8146892SSatish Balay     El::Copy(*a->emat,*b->emat);
527df311e6cSXuan Zhou   }
528df311e6cSXuan Zhou   Be->assembled = PETSC_TRUE;
529df311e6cSXuan Zhou   PetscFunctionReturn(0);
530df311e6cSXuan Zhou }
531df311e6cSXuan Zhou 
532d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
533d6223691SXuan Zhou {
534ec8cb81fSBarry Smith   Mat            Be = *B;
5355262d616SXuan Zhou   PetscErrorCode ierr;
536ce94432eSBarry Smith   MPI_Comm       comm;
5374fe7bbcaSHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
538d6223691SXuan Zhou 
539d6223691SXuan Zhou   PetscFunctionBegin;
540ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5415cb544a0SHong Zhang   /* Only out-of-place supported */
5425262d616SXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5435262d616SXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5445262d616SXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5455262d616SXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5465262d616SXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5475262d616SXuan Zhou     *B = Be;
5485262d616SXuan Zhou   }
5494fe7bbcaSHong Zhang   b = (Mat_Elemental*)Be->data;
550e8146892SSatish Balay   El::Transpose(*a->emat,*b->emat);
5515262d616SXuan Zhou   Be->assembled = PETSC_TRUE;
552d6223691SXuan Zhou   PetscFunctionReturn(0);
553d6223691SXuan Zhou }
554d6223691SXuan Zhou 
555dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A)
556dfcb0403SXuan Zhou {
557dfcb0403SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
558dfcb0403SXuan Zhou 
559dfcb0403SXuan Zhou   PetscFunctionBegin;
560e8146892SSatish Balay   El::Conjugate(*a->emat);
561dfcb0403SXuan Zhou   PetscFunctionReturn(0);
562dfcb0403SXuan Zhou }
563dfcb0403SXuan Zhou 
5644a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
5654a29722dSXuan Zhou {
566ec8cb81fSBarry Smith   Mat            Be = *B;
5674a29722dSXuan Zhou   PetscErrorCode ierr;
568ce94432eSBarry Smith   MPI_Comm       comm;
5694a29722dSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
5704a29722dSXuan Zhou 
5714a29722dSXuan Zhou   PetscFunctionBegin;
572ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5735cb544a0SHong Zhang   /* Only out-of-place supported */
5744a29722dSXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5754a29722dSXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5764a29722dSXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5774a29722dSXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5784a29722dSXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5794a29722dSXuan Zhou     *B = Be;
5804a29722dSXuan Zhou   }
5814a29722dSXuan Zhou   b = (Mat_Elemental*)Be->data;
582e8146892SSatish Balay   El::Adjoint(*a->emat,*b->emat);
5834a29722dSXuan Zhou   Be->assembled = PETSC_TRUE;
5844a29722dSXuan Zhou   PetscFunctionReturn(0);
5854a29722dSXuan Zhou }
5864a29722dSXuan Zhou 
5871f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
5881f881ff8SXuan Zhou {
5891f881ff8SXuan Zhou   Mat_Elemental     *a = (Mat_Elemental*)A->data;
5901f881ff8SXuan Zhou   PetscErrorCode    ierr;
591df311e6cSXuan Zhou   PetscElemScalar   *x;
592ea394475SSatish Balay   PetscInt          pivoting = a->pivoting;
5931f881ff8SXuan Zhou 
5941f881ff8SXuan Zhou   PetscFunctionBegin;
59545cf121fSXuan Zhou   ierr = VecCopy(B,X);CHKERRQ(ierr);
596e6dea9dbSXuan Zhou   ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
597ea394475SSatish Balay 
598e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe;
5990c18141cSBarry Smith   xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
600e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::MC,El::MR> xer(xe);
601fc54b460SXuan Zhou   switch (A->factortype) {
602fc54b460SXuan Zhou   case MAT_FACTOR_LU:
603ea394475SSatish Balay     if (pivoting == 0) {
604e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,xer);
605ea394475SSatish Balay     } else if (pivoting == 1) {
606ea394475SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,xer);
607ea394475SSatish Balay     } else { /* pivoting == 2 */
608ea394475SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,xer);
6091f881ff8SXuan Zhou     }
610fc54b460SXuan Zhou     break;
611fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
612e8146892SSatish Balay     El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,xer);
613fc54b460SXuan Zhou     break;
614fc54b460SXuan Zhou   default:
6154fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
616fc54b460SXuan Zhou     break;
6171f881ff8SXuan Zhou   }
618ea394475SSatish Balay   El::Copy(xer,xe);
619ea394475SSatish Balay 
620e6dea9dbSXuan Zhou   ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
6211f881ff8SXuan Zhou   PetscFunctionReturn(0);
6221f881ff8SXuan Zhou }
6231f881ff8SXuan Zhou 
624df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
625df311e6cSXuan Zhou {
626df311e6cSXuan Zhou   PetscErrorCode    ierr;
627df311e6cSXuan Zhou 
628df311e6cSXuan Zhou   PetscFunctionBegin;
629df311e6cSXuan Zhou   ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr);
6303d7f40dbSXuan Zhou   ierr = VecAXPY(X,1,Y);CHKERRQ(ierr);
631df311e6cSXuan Zhou   PetscFunctionReturn(0);
632df311e6cSXuan Zhou }
633df311e6cSXuan Zhou 
634ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
635ae844d54SHong Zhang {
6361f0e42cfSHong Zhang   Mat_Elemental *a=(Mat_Elemental*)A->data;
637d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
6381f0e42cfSHong Zhang   Mat_Elemental *x=(Mat_Elemental*)X->data;
639ea394475SSatish Balay   PetscInt      pivoting = a->pivoting;
6401f0e42cfSHong Zhang 
641ae844d54SHong Zhang   PetscFunctionBegin;
642e8146892SSatish Balay   El::Copy(*b->emat,*x->emat);
643fc54b460SXuan Zhou   switch (A->factortype) {
644fc54b460SXuan Zhou   case MAT_FACTOR_LU:
645ea394475SSatish Balay     if (pivoting == 0) {
646e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*x->emat);
647ea394475SSatish Balay     } else if (pivoting == 1) {
648ea394475SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*x->emat);
649ea394475SSatish Balay     } else {
650ea394475SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,*x->emat);
651d6223691SXuan Zhou     }
652fc54b460SXuan Zhou     break;
653fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
654e8146892SSatish Balay     El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,*x->emat);
655fc54b460SXuan Zhou     break;
656fc54b460SXuan Zhou   default:
6574fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
658fc54b460SXuan Zhou     break;
659fc54b460SXuan Zhou   }
660ae844d54SHong Zhang   PetscFunctionReturn(0);
661ae844d54SHong Zhang }
662ae844d54SHong Zhang 
663ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
664ae844d54SHong Zhang {
6657c920d81SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
666124af5d2SSatish Balay   PetscErrorCode ierr;
667ea394475SSatish Balay   PetscInt       pivoting = a->pivoting;
6687c920d81SXuan Zhou 
669ae844d54SHong Zhang   PetscFunctionBegin;
670ea394475SSatish Balay   if (pivoting == 0) {
671e8146892SSatish Balay     El::LU(*a->emat);
672ea394475SSatish Balay   } else if (pivoting == 1) {
673ea394475SSatish Balay     El::LU(*a->emat,*a->P);
674ea394475SSatish Balay   } else {
675ea394475SSatish Balay     El::LU(*a->emat,*a->P,*a->Q);
676d6223691SXuan Zhou   }
6771293973dSHong Zhang   A->factortype = MAT_FACTOR_LU;
678834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
679f6224b95SHong Zhang 
680f6224b95SHong Zhang   ierr = PetscFree(A->solvertype);CHKERRQ(ierr);
681f6224b95SHong Zhang   ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);CHKERRQ(ierr);
682ae844d54SHong Zhang   PetscFunctionReturn(0);
683ae844d54SHong Zhang }
684ae844d54SHong Zhang 
685d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
686d7c3f9d8SHong Zhang {
687d7c3f9d8SHong Zhang   PetscErrorCode ierr;
688d7c3f9d8SHong Zhang 
689d7c3f9d8SHong Zhang   PetscFunctionBegin;
690d7c3f9d8SHong Zhang   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
691d7c3f9d8SHong Zhang   ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr);
692d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
693d7c3f9d8SHong Zhang }
694d7c3f9d8SHong Zhang 
695d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
696d7c3f9d8SHong Zhang {
697d7c3f9d8SHong Zhang   PetscFunctionBegin;
698d7c3f9d8SHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
699d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
700d7c3f9d8SHong Zhang }
701d7c3f9d8SHong Zhang 
70245cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
70345cf121fSXuan Zhou {
70445cf121fSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
705e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::MC,El::STAR> d;
706124af5d2SSatish Balay   PetscErrorCode ierr;
70745cf121fSXuan Zhou 
70845cf121fSXuan Zhou   PetscFunctionBegin;
709e8146892SSatish Balay   El::Cholesky(El::UPPER,*a->emat);
710fc54b460SXuan Zhou   A->factortype = MAT_FACTOR_CHOLESKY;
711834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
712f6224b95SHong Zhang 
713f6224b95SHong Zhang   ierr = PetscFree(A->solvertype);CHKERRQ(ierr);
714f6224b95SHong Zhang   ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);CHKERRQ(ierr);
71545cf121fSXuan Zhou   PetscFunctionReturn(0);
71645cf121fSXuan Zhou }
71745cf121fSXuan Zhou 
71879673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
71979673f7bSHong Zhang {
720cb76c1d8SXuan Zhou   PetscErrorCode ierr;
721cb76c1d8SXuan Zhou 
722cb76c1d8SXuan Zhou   PetscFunctionBegin;
723cb76c1d8SXuan Zhou   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
724cb76c1d8SXuan Zhou   ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr);
725cb76c1d8SXuan Zhou   PetscFunctionReturn(0);
72679673f7bSHong Zhang }
727cb76c1d8SXuan Zhou 
72879673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
72979673f7bSHong Zhang {
73079673f7bSHong Zhang   PetscFunctionBegin;
73179673f7bSHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
73279673f7bSHong Zhang   PetscFunctionReturn(0);
73379673f7bSHong Zhang }
73479673f7bSHong Zhang 
73515767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type)
73615767789SHong Zhang {
73715767789SHong Zhang   PetscFunctionBegin;
73815767789SHong Zhang   *type = MATSOLVERELEMENTAL;
73915767789SHong Zhang   PetscFunctionReturn(0);
74015767789SHong Zhang }
74115767789SHong Zhang 
74215767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
7431293973dSHong Zhang {
7441293973dSHong Zhang   Mat            B;
7451293973dSHong Zhang   PetscErrorCode ierr;
7461293973dSHong Zhang 
7471293973dSHong Zhang   PetscFunctionBegin;
7481293973dSHong Zhang   /* Create the factorization matrix */
749ce94432eSBarry Smith   ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
7501293973dSHong Zhang   ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
7511293973dSHong Zhang   ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr);
7521293973dSHong Zhang   ierr = MatSetUp(B);CHKERRQ(ierr);
7531293973dSHong Zhang   B->factortype = ftype;
75400c67f3bSHong Zhang   ierr = PetscFree(B->solvertype);CHKERRQ(ierr);
75500c67f3bSHong Zhang   ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&B->solvertype);CHKERRQ(ierr);
75600c67f3bSHong Zhang 
757bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr);
7581293973dSHong Zhang   *F            = B;
7591293973dSHong Zhang   PetscFunctionReturn(0);
7601293973dSHong Zhang }
7611293973dSHong Zhang 
76229b38603SBarry Smith PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_Elemental(void)
76342c9c57cSBarry Smith {
76418713533SBarry Smith   PetscErrorCode ierr;
76518713533SBarry Smith 
76642c9c57cSBarry Smith   PetscFunctionBegin;
76742c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_LU,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
76842c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
76942c9c57cSBarry Smith   PetscFunctionReturn(0);
77042c9c57cSBarry Smith }
77142c9c57cSBarry Smith 
7721f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
7731f881ff8SXuan Zhou {
7741f881ff8SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7751f881ff8SXuan Zhou 
7761f881ff8SXuan Zhou   PetscFunctionBegin;
7771f881ff8SXuan Zhou   switch (type){
7781f881ff8SXuan Zhou   case NORM_1:
779e8146892SSatish Balay     *nrm = El::OneNorm(*a->emat);
7801f881ff8SXuan Zhou     break;
7811f881ff8SXuan Zhou   case NORM_FROBENIUS:
782e8146892SSatish Balay     *nrm = El::FrobeniusNorm(*a->emat);
7831f881ff8SXuan Zhou     break;
7841f881ff8SXuan Zhou   case NORM_INFINITY:
785e8146892SSatish Balay     *nrm = El::InfinityNorm(*a->emat);
7861f881ff8SXuan Zhou     break;
7871f881ff8SXuan Zhou   default:
7881f881ff8SXuan Zhou     printf("Error: unsupported norm type!\n");
7891f881ff8SXuan Zhou   }
7901f881ff8SXuan Zhou   PetscFunctionReturn(0);
7911f881ff8SXuan Zhou }
7921f881ff8SXuan Zhou 
7935262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A)
7945262d616SXuan Zhou {
7955262d616SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7965262d616SXuan Zhou 
7975262d616SXuan Zhou   PetscFunctionBegin;
798e8146892SSatish Balay   El::Zero(*a->emat);
7995262d616SXuan Zhou   PetscFunctionReturn(0);
8005262d616SXuan Zhou }
8015262d616SXuan Zhou 
802db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
803db31f6deSJed Brown {
804db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
805db31f6deSJed Brown   PetscErrorCode ierr;
806db31f6deSJed Brown   PetscInt       i,m,shift,stride,*idx;
807db31f6deSJed Brown 
808db31f6deSJed Brown   PetscFunctionBegin;
809db31f6deSJed Brown   if (rows) {
810db31f6deSJed Brown     m = a->emat->LocalHeight();
811db31f6deSJed Brown     shift = a->emat->ColShift();
812db31f6deSJed Brown     stride = a->emat->ColStride();
813785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
814db31f6deSJed Brown     for (i=0; i<m; i++) {
815db31f6deSJed Brown       PetscInt rank,offset;
816db31f6deSJed Brown       E2RO(A,0,shift+i*stride,&rank,&offset);
817db31f6deSJed Brown       RO2P(A,0,rank,offset,&idx[i]);
818db31f6deSJed Brown     }
819db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
820db31f6deSJed Brown   }
821db31f6deSJed Brown   if (cols) {
822db31f6deSJed Brown     m = a->emat->LocalWidth();
823db31f6deSJed Brown     shift = a->emat->RowShift();
824db31f6deSJed Brown     stride = a->emat->RowStride();
825785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
826db31f6deSJed Brown     for (i=0; i<m; i++) {
827db31f6deSJed Brown       PetscInt rank,offset;
828db31f6deSJed Brown       E2RO(A,1,shift+i*stride,&rank,&offset);
829db31f6deSJed Brown       RO2P(A,1,rank,offset,&idx[i]);
830db31f6deSJed Brown     }
831db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr);
832db31f6deSJed Brown   }
833db31f6deSJed Brown   PetscFunctionReturn(0);
834db31f6deSJed Brown }
835db31f6deSJed Brown 
83619fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
837af295397SXuan Zhou {
8382ef0cf24SXuan Zhou   Mat                Bmpi;
839af295397SXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
840ce94432eSBarry Smith   MPI_Comm           comm;
8412ef0cf24SXuan Zhou   PetscErrorCode     ierr;
842fa9e26c8SHong Zhang   IS                 isrows,iscols;
843fa9e26c8SHong Zhang   PetscInt           rrank,ridx,crank,cidx,nrows,ncols,i,j,erow,ecol,elrow,elcol;
844fa9e26c8SHong Zhang   const PetscInt     *rows,*cols;
845df311e6cSXuan Zhou   PetscElemScalar    v;
846fa9e26c8SHong Zhang   const El::Grid     &grid = a->emat->Grid();
847af295397SXuan Zhou 
848af295397SXuan Zhou   PetscFunctionBegin;
849ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
850a000e53bSHong Zhang 
851de0a22f0SHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
852de0a22f0SHong Zhang     Bmpi = *B;
853de0a22f0SHong Zhang   } else {
854af295397SXuan Zhou     ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr);
855af295397SXuan Zhou     ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
8562ef0cf24SXuan Zhou     ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr);
857af295397SXuan Zhou     ierr = MatSetUp(Bmpi);CHKERRQ(ierr);
858de0a22f0SHong Zhang   }
859fa9e26c8SHong Zhang 
860fa9e26c8SHong Zhang   /* Get local entries of A */
861fa9e26c8SHong Zhang   ierr = MatGetOwnershipIS(A,&isrows,&iscols);CHKERRQ(ierr);
862fa9e26c8SHong Zhang   ierr = ISGetLocalSize(isrows,&nrows);CHKERRQ(ierr);
863fa9e26c8SHong Zhang   ierr = ISGetIndices(isrows,&rows);CHKERRQ(ierr);
864fa9e26c8SHong Zhang   ierr = ISGetLocalSize(iscols,&ncols);CHKERRQ(ierr);
865fa9e26c8SHong Zhang   ierr = ISGetIndices(iscols,&cols);CHKERRQ(ierr);
866fa9e26c8SHong Zhang 
86757299f2fSHong Zhang   if (a->roworiented) {
8682ef0cf24SXuan Zhou     for (i=0; i<nrows; i++) {
869fa9e26c8SHong Zhang       P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */
8702ef0cf24SXuan Zhou       RO2E(A,0,rrank,ridx,&erow);
8712ef0cf24SXuan Zhou       if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
8722ef0cf24SXuan Zhou       for (j=0; j<ncols; j++) {
873fa9e26c8SHong Zhang         P2RO(A,1,cols[j],&crank,&cidx);
8742ef0cf24SXuan Zhou         RO2E(A,1,crank,cidx,&ecol);
8752ef0cf24SXuan Zhou         if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
876fa9e26c8SHong Zhang 
877fa9e26c8SHong Zhang         elrow = erow / grid.MCSize(); /* Elemental local row index */
878fa9e26c8SHong Zhang         elcol = ecol / grid.MRSize(); /* Elemental local column index */
879fa9e26c8SHong Zhang         v = a->emat->GetLocal(elrow,elcol);
880fa9e26c8SHong Zhang         ierr = MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
8812ef0cf24SXuan Zhou       }
8822ef0cf24SXuan Zhou     }
88357299f2fSHong Zhang   } else { /* column-oriented */
88457299f2fSHong Zhang     for (j=0; j<ncols; j++) {
88557299f2fSHong Zhang       P2RO(A,1,cols[j],&crank,&cidx);
88657299f2fSHong Zhang       RO2E(A,1,crank,cidx,&ecol);
88757299f2fSHong Zhang       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
88857299f2fSHong Zhang       for (i=0; i<nrows; i++) {
88957299f2fSHong Zhang         P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */
89057299f2fSHong Zhang         RO2E(A,0,rrank,ridx,&erow);
89157299f2fSHong Zhang         if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
89257299f2fSHong Zhang 
89357299f2fSHong Zhang         elrow = erow / grid.MCSize(); /* Elemental local row index */
89457299f2fSHong Zhang         elcol = ecol / grid.MRSize(); /* Elemental local column index */
89557299f2fSHong Zhang         v = a->emat->GetLocal(elrow,elcol);
89657299f2fSHong Zhang         ierr = MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
89757299f2fSHong Zhang       }
89857299f2fSHong Zhang     }
89957299f2fSHong Zhang   }
900af295397SXuan Zhou   ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
901af295397SXuan Zhou   ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
902511c6705SHong Zhang   if (reuse == MAT_INPLACE_MATRIX) {
90328be2f97SBarry Smith     ierr = MatHeaderReplace(A,&Bmpi);CHKERRQ(ierr);
904c4ad791aSXuan Zhou   } else {
905c4ad791aSXuan Zhou     *B = Bmpi;
906c4ad791aSXuan Zhou   }
907fa9e26c8SHong Zhang   ierr = ISDestroy(&isrows);CHKERRQ(ierr);
908fa9e26c8SHong Zhang   ierr = ISDestroy(&iscols);CHKERRQ(ierr);
909af295397SXuan Zhou   PetscFunctionReturn(0);
910af295397SXuan Zhou }
911af295397SXuan Zhou 
912cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
913af8000cdSHong Zhang {
914af8000cdSHong Zhang   Mat               mat_elemental;
915af8000cdSHong Zhang   PetscErrorCode    ierr;
916af8000cdSHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols;
917af8000cdSHong Zhang   const PetscInt    *cols;
918af8000cdSHong Zhang   const PetscScalar *vals;
919af8000cdSHong Zhang 
920af8000cdSHong Zhang   PetscFunctionBegin;
921af8000cdSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
922af8000cdSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
923af8000cdSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
924af8000cdSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
925af8000cdSHong Zhang   for (row=0; row<M; row++) {
926af8000cdSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
927af8000cdSHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
928af8000cdSHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
929af8000cdSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
930af8000cdSHong Zhang   }
931af8000cdSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
932af8000cdSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
933af8000cdSHong Zhang 
934511c6705SHong Zhang   if (reuse == MAT_INPLACE_MATRIX) {
93528be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
936af8000cdSHong Zhang   } else {
937af8000cdSHong Zhang     *newmat = mat_elemental;
938af8000cdSHong Zhang   }
939af8000cdSHong Zhang   PetscFunctionReturn(0);
940af8000cdSHong Zhang }
941af8000cdSHong Zhang 
942cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9435d7652ecSHong Zhang {
9445d7652ecSHong Zhang   Mat               mat_elemental;
9455d7652ecSHong Zhang   PetscErrorCode    ierr;
9465d7652ecSHong Zhang   PetscInt          row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j;
9475d7652ecSHong Zhang   const PetscInt    *cols;
9485d7652ecSHong Zhang   const PetscScalar *vals;
9495d7652ecSHong Zhang 
9505d7652ecSHong Zhang   PetscFunctionBegin;
9515d7652ecSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9525d7652ecSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);CHKERRQ(ierr);
9535d7652ecSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9545d7652ecSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9555d7652ecSHong Zhang   for (row=rstart; row<rend; row++) {
9565d7652ecSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9575d7652ecSHong Zhang     for (j=0; j<ncols; j++) {
9585d7652ecSHong Zhang       /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9595d7652ecSHong Zhang       ierr = MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);CHKERRQ(ierr);
9605d7652ecSHong Zhang     }
9615d7652ecSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9625d7652ecSHong Zhang   }
9635d7652ecSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9645d7652ecSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9655d7652ecSHong Zhang 
966511c6705SHong Zhang   if (reuse == MAT_INPLACE_MATRIX) {
96728be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
9685d7652ecSHong Zhang   } else {
9695d7652ecSHong Zhang     *newmat = mat_elemental;
9705d7652ecSHong Zhang   }
9715d7652ecSHong Zhang   PetscFunctionReturn(0);
9725d7652ecSHong Zhang }
9735d7652ecSHong Zhang 
974cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9756214f412SHong Zhang {
9766214f412SHong Zhang   Mat               mat_elemental;
9776214f412SHong Zhang   PetscErrorCode    ierr;
9786214f412SHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols,j;
9796214f412SHong Zhang   const PetscInt    *cols;
9806214f412SHong Zhang   const PetscScalar *vals;
9816214f412SHong Zhang 
9826214f412SHong Zhang   PetscFunctionBegin;
9836214f412SHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9846214f412SHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
9856214f412SHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9866214f412SHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9876214f412SHong Zhang   ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr);
9886214f412SHong Zhang   for (row=0; row<M; row++) {
9896214f412SHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9906214f412SHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9916214f412SHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
9926214f412SHong Zhang     for (j=0; j<ncols; j++) { /* lower triangular part */
9936214f412SHong Zhang       if (cols[j] == row) continue;
9946214f412SHong Zhang       ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr);
9956214f412SHong Zhang     }
9966214f412SHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9976214f412SHong Zhang   }
9986214f412SHong Zhang   ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr);
9996214f412SHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
10006214f412SHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
10016214f412SHong Zhang 
1002511c6705SHong Zhang   if (reuse == MAT_INPLACE_MATRIX) {
100328be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
10046214f412SHong Zhang   } else {
10056214f412SHong Zhang     *newmat = mat_elemental;
10066214f412SHong Zhang   }
10076214f412SHong Zhang   PetscFunctionReturn(0);
10086214f412SHong Zhang }
10096214f412SHong Zhang 
1010cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_MPISBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
10116214f412SHong Zhang {
10126214f412SHong Zhang   Mat               mat_elemental;
10136214f412SHong Zhang   PetscErrorCode    ierr;
10146214f412SHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols,j,rstart=A->rmap->rstart,rend=A->rmap->rend;
10156214f412SHong Zhang   const PetscInt    *cols;
10166214f412SHong Zhang   const PetscScalar *vals;
10176214f412SHong Zhang 
10186214f412SHong Zhang   PetscFunctionBegin;
10196214f412SHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
10206214f412SHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
10216214f412SHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
10226214f412SHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
10236214f412SHong Zhang   ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr);
10246214f412SHong Zhang   for (row=rstart; row<rend; row++) {
10256214f412SHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
10266214f412SHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
10276214f412SHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
10286214f412SHong Zhang     for (j=0; j<ncols; j++) { /* lower triangular part */
10296214f412SHong Zhang       if (cols[j] == row) continue;
10306214f412SHong Zhang       ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr);
10316214f412SHong Zhang     }
10326214f412SHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
10336214f412SHong Zhang   }
10346214f412SHong Zhang   ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr);
10356214f412SHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
10366214f412SHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
10376214f412SHong Zhang 
1038511c6705SHong Zhang   if (reuse == MAT_INPLACE_MATRIX) {
103928be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
10406214f412SHong Zhang   } else {
10416214f412SHong Zhang     *newmat = mat_elemental;
10426214f412SHong Zhang   }
10436214f412SHong Zhang   PetscFunctionReturn(0);
10446214f412SHong Zhang }
10456214f412SHong Zhang 
1046db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A)
1047db31f6deSJed Brown {
1048db31f6deSJed Brown   Mat_Elemental      *a = (Mat_Elemental*)A->data;
1049db31f6deSJed Brown   PetscErrorCode     ierr;
10505e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
10515e9f5b67SHong Zhang   PetscBool          flg;
10525e9f5b67SHong Zhang   MPI_Comm           icomm;
1053db31f6deSJed Brown 
1054db31f6deSJed Brown   PetscFunctionBegin;
1055db31f6deSJed Brown   delete a->emat;
1056ea394475SSatish Balay   delete a->P;
1057ea394475SSatish Balay   delete a->Q;
10585e9f5b67SHong Zhang 
1059e8146892SSatish Balay   El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
10600c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
10615e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
10625e9f5b67SHong Zhang   if (--commgrid->grid_refct == 0) {
10635e9f5b67SHong Zhang     delete commgrid->grid;
10645e9f5b67SHong Zhang     ierr = PetscFree(commgrid);CHKERRQ(ierr);
1065a000e53bSHong Zhang     ierr = MPI_Keyval_free(&Petsc_Elemental_keyval);CHKERRQ(ierr);
10665e9f5b67SHong Zhang   }
10675e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
1068bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr);
1069bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr);
1070db31f6deSJed Brown   ierr = PetscFree(A->data);CHKERRQ(ierr);
1071db31f6deSJed Brown   PetscFunctionReturn(0);
1072db31f6deSJed Brown }
1073db31f6deSJed Brown 
1074db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A)
1075db31f6deSJed Brown {
1076db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
1077db31f6deSJed Brown   PetscErrorCode ierr;
1078db31f6deSJed Brown   PetscMPIInt    rsize,csize;
1079db31f6deSJed Brown 
1080db31f6deSJed Brown   PetscFunctionBegin;
1081db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
1082db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
1083db31f6deSJed Brown 
1084efb79153SJack Poulson   a->emat->Resize(A->rmap->N,A->cmap->N);CHKERRQ(ierr);
1085e8146892SSatish Balay   El::Zero(*a->emat);
1086db31f6deSJed Brown 
1087db31f6deSJed Brown   ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr);
1088db31f6deSJed Brown   ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr);
1089ce94432eSBarry Smith   if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
1090db31f6deSJed Brown   a->commsize = rsize;
1091db31f6deSJed Brown   a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
1092db31f6deSJed Brown   a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
1093db31f6deSJed Brown   a->m[0]  = A->rmap->N / rsize + (a->mr[0] != rsize);
1094db31f6deSJed Brown   a->m[1]  = A->cmap->N / csize + (a->mr[1] != csize);
1095db31f6deSJed Brown   PetscFunctionReturn(0);
1096db31f6deSJed Brown }
1097db31f6deSJed Brown 
1098aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
1099aae2c449SHong Zhang {
1100aae2c449SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
1101aae2c449SHong Zhang 
1102aae2c449SHong Zhang   PetscFunctionBegin;
1103fbbcb730SJack Poulson   /* printf("Calling ProcessQueues\n"); */
1104fbbcb730SJack Poulson   a->emat->ProcessQueues();
1105fbbcb730SJack Poulson   /* printf("Finished ProcessQueues\n"); */
1106aae2c449SHong Zhang   PetscFunctionReturn(0);
1107aae2c449SHong Zhang }
1108aae2c449SHong Zhang 
1109aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
1110aae2c449SHong Zhang {
1111aae2c449SHong Zhang   PetscFunctionBegin;
1112aae2c449SHong Zhang   /* Currently does nothing */
1113aae2c449SHong Zhang   PetscFunctionReturn(0);
1114aae2c449SHong Zhang }
1115aae2c449SHong Zhang 
11167b30e0f1SHong Zhang PetscErrorCode MatLoad_Elemental(Mat newMat, PetscViewer viewer)
11177b30e0f1SHong Zhang {
11187b30e0f1SHong Zhang   PetscErrorCode ierr;
11197b30e0f1SHong Zhang   Mat            Adense,Ae;
11207b30e0f1SHong Zhang   MPI_Comm       comm;
11217b30e0f1SHong Zhang 
11227b30e0f1SHong Zhang   PetscFunctionBegin;
11237b30e0f1SHong Zhang   ierr = PetscObjectGetComm((PetscObject)newMat,&comm);CHKERRQ(ierr);
11247b30e0f1SHong Zhang   ierr = MatCreate(comm,&Adense);CHKERRQ(ierr);
11257b30e0f1SHong Zhang   ierr = MatSetType(Adense,MATDENSE);CHKERRQ(ierr);
11267b30e0f1SHong Zhang   ierr = MatLoad(Adense,viewer);CHKERRQ(ierr);
11277b30e0f1SHong Zhang   ierr = MatConvert(Adense, MATELEMENTAL, MAT_INITIAL_MATRIX,&Ae);CHKERRQ(ierr);
11287b30e0f1SHong Zhang   ierr = MatDestroy(&Adense);CHKERRQ(ierr);
112928be2f97SBarry Smith   ierr = MatHeaderReplace(newMat,&Ae);CHKERRQ(ierr);
11307b30e0f1SHong Zhang   PetscFunctionReturn(0);
11317b30e0f1SHong Zhang }
11327b30e0f1SHong Zhang 
113340d92e34SHong Zhang /* -------------------------------------------------------------------*/
113440d92e34SHong Zhang static struct _MatOps MatOps_Values = {
113540d92e34SHong Zhang        MatSetValues_Elemental,
113640d92e34SHong Zhang        0,
113740d92e34SHong Zhang        0,
113840d92e34SHong Zhang        MatMult_Elemental,
113940d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental,
11409426833fSXuan Zhou        MatMultTranspose_Elemental,
1141e883f9d5SXuan Zhou        MatMultTransposeAdd_Elemental,
114240d92e34SHong Zhang        MatSolve_Elemental,
1143df311e6cSXuan Zhou        MatSolveAdd_Elemental,
11442ef15aa2SHong Zhang        0,
11452ef15aa2SHong Zhang /*10*/ 0,
114640d92e34SHong Zhang        MatLUFactor_Elemental,
114740d92e34SHong Zhang        MatCholeskyFactor_Elemental,
114840d92e34SHong Zhang        0,
114940d92e34SHong Zhang        MatTranspose_Elemental,
115040d92e34SHong Zhang /*15*/ MatGetInfo_Elemental,
115140d92e34SHong Zhang        0,
115261119200SXuan Zhou        MatGetDiagonal_Elemental,
1153ade3cc5eSXuan Zhou        MatDiagonalScale_Elemental,
115440d92e34SHong Zhang        MatNorm_Elemental,
115540d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental,
115640d92e34SHong Zhang        MatAssemblyEnd_Elemental,
115732d7a744SHong Zhang        MatSetOption_Elemental,
115840d92e34SHong Zhang        MatZeroEntries_Elemental,
115940d92e34SHong Zhang /*24*/ 0,
116040d92e34SHong Zhang        MatLUFactorSymbolic_Elemental,
116140d92e34SHong Zhang        MatLUFactorNumeric_Elemental,
116240d92e34SHong Zhang        MatCholeskyFactorSymbolic_Elemental,
116340d92e34SHong Zhang        MatCholeskyFactorNumeric_Elemental,
116440d92e34SHong Zhang /*29*/ MatSetUp_Elemental,
116540d92e34SHong Zhang        0,
116640d92e34SHong Zhang        0,
116740d92e34SHong Zhang        0,
116840d92e34SHong Zhang        0,
1169df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental,
117040d92e34SHong Zhang        0,
117140d92e34SHong Zhang        0,
117240d92e34SHong Zhang        0,
117340d92e34SHong Zhang        0,
117440d92e34SHong Zhang /*39*/ MatAXPY_Elemental,
117540d92e34SHong Zhang        0,
117640d92e34SHong Zhang        0,
117740d92e34SHong Zhang        0,
117840d92e34SHong Zhang        MatCopy_Elemental,
117940d92e34SHong Zhang /*44*/ 0,
118040d92e34SHong Zhang        MatScale_Elemental,
11817d68702bSBarry Smith        MatShift_Basic,
118240d92e34SHong Zhang        0,
118340d92e34SHong Zhang        0,
118440d92e34SHong Zhang /*49*/ 0,
118540d92e34SHong Zhang        0,
118640d92e34SHong Zhang        0,
118740d92e34SHong Zhang        0,
118840d92e34SHong Zhang        0,
118940d92e34SHong Zhang /*54*/ 0,
119040d92e34SHong Zhang        0,
119140d92e34SHong Zhang        0,
119240d92e34SHong Zhang        0,
119340d92e34SHong Zhang        0,
119440d92e34SHong Zhang /*59*/ 0,
119540d92e34SHong Zhang        MatDestroy_Elemental,
119640d92e34SHong Zhang        MatView_Elemental,
119740d92e34SHong Zhang        0,
119840d92e34SHong Zhang        0,
119940d92e34SHong Zhang /*64*/ 0,
120040d92e34SHong Zhang        0,
120140d92e34SHong Zhang        0,
120240d92e34SHong Zhang        0,
120340d92e34SHong Zhang        0,
120440d92e34SHong Zhang /*69*/ 0,
120540d92e34SHong Zhang        0,
12062ef0cf24SXuan Zhou        MatConvert_Elemental_Dense,
120740d92e34SHong Zhang        0,
120840d92e34SHong Zhang        0,
120940d92e34SHong Zhang /*74*/ 0,
121040d92e34SHong Zhang        0,
121140d92e34SHong Zhang        0,
121240d92e34SHong Zhang        0,
121340d92e34SHong Zhang        0,
121440d92e34SHong Zhang /*79*/ 0,
121540d92e34SHong Zhang        0,
121640d92e34SHong Zhang        0,
121740d92e34SHong Zhang        0,
12187b30e0f1SHong Zhang        MatLoad_Elemental,
121940d92e34SHong Zhang /*84*/ 0,
122040d92e34SHong Zhang        0,
122140d92e34SHong Zhang        0,
122240d92e34SHong Zhang        0,
122340d92e34SHong Zhang        0,
122440d92e34SHong Zhang /*89*/ MatMatMult_Elemental,
122540d92e34SHong Zhang        MatMatMultSymbolic_Elemental,
122640d92e34SHong Zhang        MatMatMultNumeric_Elemental,
122740d92e34SHong Zhang        0,
122840d92e34SHong Zhang        0,
122940d92e34SHong Zhang /*94*/ 0,
1230df311e6cSXuan Zhou        MatMatTransposeMult_Elemental,
1231df311e6cSXuan Zhou        MatMatTransposeMultSymbolic_Elemental,
1232df311e6cSXuan Zhou        MatMatTransposeMultNumeric_Elemental,
123340d92e34SHong Zhang        0,
123440d92e34SHong Zhang /*99*/ 0,
123540d92e34SHong Zhang        0,
123640d92e34SHong Zhang        0,
1237dfcb0403SXuan Zhou        MatConjugate_Elemental,
123840d92e34SHong Zhang        0,
123940d92e34SHong Zhang /*104*/0,
124040d92e34SHong Zhang        0,
124140d92e34SHong Zhang        0,
124240d92e34SHong Zhang        0,
124340d92e34SHong Zhang        0,
124440d92e34SHong Zhang /*109*/MatMatSolve_Elemental,
124540d92e34SHong Zhang        0,
124640d92e34SHong Zhang        0,
124740d92e34SHong Zhang        0,
124840d92e34SHong Zhang        0,
124940d92e34SHong Zhang /*114*/0,
125040d92e34SHong Zhang        0,
125140d92e34SHong Zhang        0,
125240d92e34SHong Zhang        0,
125340d92e34SHong Zhang        0,
125440d92e34SHong Zhang /*119*/0,
12554a29722dSXuan Zhou        MatHermitianTranspose_Elemental,
125640d92e34SHong Zhang        0,
125740d92e34SHong Zhang        0,
125840d92e34SHong Zhang        0,
125940d92e34SHong Zhang /*124*/0,
126040d92e34SHong Zhang        0,
126140d92e34SHong Zhang        0,
126240d92e34SHong Zhang        0,
126340d92e34SHong Zhang        0,
126440d92e34SHong Zhang /*129*/0,
126540d92e34SHong Zhang        0,
126640d92e34SHong Zhang        0,
126740d92e34SHong Zhang        0,
126840d92e34SHong Zhang        0,
126940d92e34SHong Zhang /*134*/0,
127040d92e34SHong Zhang        0,
127140d92e34SHong Zhang        0,
127240d92e34SHong Zhang        0,
127340d92e34SHong Zhang        0
127440d92e34SHong Zhang };
127540d92e34SHong Zhang 
1276ed36708cSHong Zhang /*MC
1277ed36708cSHong Zhang    MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1278ed36708cSHong Zhang 
1279c2b89b5dSBarry Smith   Use ./configure --download-elemental to install PETSc to use Elemental
1280c2b89b5dSBarry Smith 
1281c2b89b5dSBarry Smith   Use -pc_type lu -pc_factor_mat_solver_package elemental to us this direct solver
1282c2b89b5dSBarry Smith 
1283ed36708cSHong Zhang    Options Database Keys:
12845cc86fc1SJed Brown + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
12855cc86fc1SJed Brown - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1286ed36708cSHong Zhang 
1287ed36708cSHong Zhang   Level: beginner
1288ed36708cSHong Zhang 
12895cb544a0SHong Zhang .seealso: MATDENSE
1290ed36708cSHong Zhang M*/
12914a29722dSXuan Zhou 
12928cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1293db31f6deSJed Brown {
1294db31f6deSJed Brown   Mat_Elemental      *a;
1295db31f6deSJed Brown   PetscErrorCode     ierr;
12965682a260SJack Poulson   PetscBool          flg,flg1;
12975e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
12985e9f5b67SHong Zhang   MPI_Comm           icomm;
12995682a260SJack Poulson   PetscInt           optv1;
1300db31f6deSJed Brown 
1301db31f6deSJed Brown   PetscFunctionBegin;
1302607a6623SBarry Smith   ierr = PetscElementalInitializePackage();CHKERRQ(ierr);
130340d92e34SHong Zhang   ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
130440d92e34SHong Zhang   A->insertmode = NOT_SET_VALUES;
1305db31f6deSJed Brown 
1306b00a9115SJed Brown   ierr = PetscNewLog(A,&a);CHKERRQ(ierr);
1307db31f6deSJed Brown   A->data = (void*)a;
1308db31f6deSJed Brown 
1309db31f6deSJed Brown   /* Set up the elemental matrix */
1310e8146892SSatish Balay   El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
13115e9f5b67SHong Zhang 
13125e9f5b67SHong Zhang   /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
13135e9f5b67SHong Zhang   if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
13142a808120SBarry Smith     ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);CHKERRQ(ierr);
1315a000e53bSHong Zhang     /* ierr = MPI_Comm_create_Keyval(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0); -- new version? */
13165e9f5b67SHong Zhang   }
13170c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
13185e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
13195e9f5b67SHong Zhang   if (!flg) {
1320b00a9115SJed Brown     ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr);
13215cb544a0SHong Zhang 
1322ce94432eSBarry Smith     ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr);
1323e8146892SSatish Balay     /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until El::Grid() is called */
1324e8146892SSatish Balay     ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",El::mpi::Size(cxxcomm),&optv1,&flg1);CHKERRQ(ierr);
13255682a260SJack Poulson     if (flg1) {
1326e8146892SSatish Balay       if (El::mpi::Size(cxxcomm) % optv1 != 0) {
1327e8146892SSatish Balay         SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)El::mpi::Size(cxxcomm));
1328ed667823SXuan Zhou       }
1329e8146892SSatish Balay       commgrid->grid = new El::Grid(cxxcomm,optv1); /* use user-provided grid height */
13302ef0cf24SXuan Zhou     } else {
1331e8146892SSatish Balay       commgrid->grid = new El::Grid(cxxcomm); /* use Elemental default grid sizes */
1332da0640a4SHong Zhang       /* printf("new commgrid->grid = %p\n",commgrid->grid);  -- memory leak revealed by valgrind? */
1333ed667823SXuan Zhou     }
13345e9f5b67SHong Zhang     commgrid->grid_refct = 1;
13355e9f5b67SHong Zhang     ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr);
1336ea394475SSatish Balay 
1337ea394475SSatish Balay     a->pivoting    = 1;
1338ea394475SSatish Balay     ierr = PetscOptionsInt("-mat_elemental_pivoting","Pivoting","None",a->pivoting,&a->pivoting,NULL);CHKERRQ(ierr);
1339ea394475SSatish Balay 
13402a808120SBarry Smith     ierr = PetscOptionsEnd();CHKERRQ(ierr);
13415e9f5b67SHong Zhang   } else {
13425e9f5b67SHong Zhang     commgrid->grid_refct++;
13435e9f5b67SHong Zhang   }
13445e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
13455e9f5b67SHong Zhang   a->grid        = commgrid->grid;
1346e8146892SSatish Balay   a->emat        = new El::DistMatrix<PetscElemScalar>(*a->grid);
134732d7a744SHong Zhang   a->roworiented = PETSC_TRUE;
1348db31f6deSJed Brown 
1349bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr);
1350db31f6deSJed Brown   ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr);
1351db31f6deSJed Brown   PetscFunctionReturn(0);
1352db31f6deSJed Brown }
1353