xref: /petsc/src/mat/impls/elemental/matelem.cxx (revision b00a91154f763f12aa55f3d53a3f2776f15f49e3)
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 #undef __FUNCT__
10db31f6deSJed Brown #define __FUNCT__ "PetscElementalInitializePackage"
11db31f6deSJed Brown /*@C
12db31f6deSJed Brown    PetscElementalInitializePackage - Initialize Elemental package
13db31f6deSJed Brown 
14db31f6deSJed Brown    Logically Collective
15db31f6deSJed Brown 
16db31f6deSJed Brown    Level: developer
17db31f6deSJed Brown 
18db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalFinalizePackage()
19db31f6deSJed Brown @*/
20607a6623SBarry Smith PetscErrorCode PetscElementalInitializePackage(void)
21db31f6deSJed Brown {
22db31f6deSJed Brown   PetscErrorCode ierr;
23db31f6deSJed Brown 
24db31f6deSJed Brown   PetscFunctionBegin;
25db31f6deSJed Brown   if (elem::Initialized()) PetscFunctionReturn(0);
26db31f6deSJed Brown   { /* 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 */
27db31f6deSJed Brown     int zero = 0;
28db31f6deSJed Brown     char **nothing = 0;
29db31f6deSJed Brown     elem::Initialize(zero,nothing);
30db31f6deSJed Brown   }
31db31f6deSJed Brown   ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr);
32db31f6deSJed Brown   PetscFunctionReturn(0);
33db31f6deSJed Brown }
34db31f6deSJed Brown 
35db31f6deSJed Brown #undef __FUNCT__
36db31f6deSJed Brown #define __FUNCT__ "PetscElementalFinalizePackage"
37db31f6deSJed Brown /*@C
38db31f6deSJed Brown    PetscElementalFinalizePackage - Finalize Elemental package
39db31f6deSJed Brown 
40db31f6deSJed Brown    Logically Collective
41db31f6deSJed Brown 
42db31f6deSJed Brown    Level: developer
43db31f6deSJed Brown 
44db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage()
45db31f6deSJed Brown @*/
46db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void)
47db31f6deSJed Brown {
48db31f6deSJed Brown 
49db31f6deSJed Brown   PetscFunctionBegin;
50db31f6deSJed Brown   elem::Finalize();
51db31f6deSJed Brown   PetscFunctionReturn(0);
52db31f6deSJed Brown }
53db31f6deSJed Brown 
54db31f6deSJed Brown #undef __FUNCT__
55db31f6deSJed Brown #define __FUNCT__ "MatView_Elemental"
56db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
57db31f6deSJed Brown {
58db31f6deSJed Brown   PetscErrorCode ierr;
59db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
60db31f6deSJed Brown   PetscBool      iascii;
61db31f6deSJed Brown 
62db31f6deSJed Brown   PetscFunctionBegin;
63db31f6deSJed Brown   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
64db31f6deSJed Brown   if (iascii) {
65db31f6deSJed Brown     PetscViewerFormat format;
66db31f6deSJed Brown     ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
67db31f6deSJed Brown     if (format == PETSC_VIEWER_ASCII_INFO) {
6879673f7bSHong Zhang       /* call elemental viewing function */
692d8adcc7SHong Zhang       ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr);
70ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr);
71ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr);
724fe7bbcaSHong Zhang       if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
7379673f7bSHong Zhang         /* call elemental viewing function */
74ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr);
754fe7bbcaSHong Zhang       }
7679673f7bSHong Zhang 
77db31f6deSJed Brown     } else if (format == PETSC_VIEWER_DEFAULT) {
78db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr);
79dae58748SBarry Smith       ierr = PetscObjectPrintClassNamePrefixType((PetscObject)A,viewer);CHKERRQ(ierr);
807a583510SJack Poulson       elem::Print( *a->emat, "Elemental matrix (cyclic ordering)" );
81db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr);
82834d3fecSHong Zhang       if (A->factortype == MAT_FACTOR_NONE){
8361119200SXuan Zhou         Mat Adense;
84ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
8561119200SXuan Zhou         ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
8661119200SXuan Zhou         ierr = MatView(Adense,viewer);CHKERRQ(ierr);
8761119200SXuan Zhou         ierr = MatDestroy(&Adense);CHKERRQ(ierr);
88834d3fecSHong Zhang       }
89ce94432eSBarry Smith     } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
90d2daa67eSHong Zhang   } else {
915cb544a0SHong Zhang     /* convert to dense format and call MatView() */
9261119200SXuan Zhou     Mat Adense;
93ce94432eSBarry Smith     ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
9461119200SXuan Zhou     ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
9561119200SXuan Zhou     ierr = MatView(Adense,viewer);CHKERRQ(ierr);
9661119200SXuan Zhou     ierr = MatDestroy(&Adense);CHKERRQ(ierr);
97d2daa67eSHong Zhang   }
98db31f6deSJed Brown   PetscFunctionReturn(0);
99db31f6deSJed Brown }
100db31f6deSJed Brown 
101db31f6deSJed Brown #undef __FUNCT__
102180a43e4SHong Zhang #define __FUNCT__ "MatGetInfo_Elemental"
10315767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
104180a43e4SHong Zhang {
10515767789SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
10615767789SHong Zhang   PetscMPIInt    rank;
10715767789SHong Zhang 
108180a43e4SHong Zhang   PetscFunctionBegin;
109ce94432eSBarry Smith   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
11015767789SHong Zhang 
11115767789SHong Zhang   /* if (!rank) printf("          .........MatGetInfo_Elemental ...\n"); */
1125cb544a0SHong Zhang   info->block_size     = 1.0;
11315767789SHong Zhang 
11415767789SHong Zhang   if (flag == MAT_LOCAL) {
11515767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
11615767789SHong Zhang     info->nz_used        = info->nz_allocated;
11715767789SHong Zhang   } else if (flag == MAT_GLOBAL_MAX) {
118ce94432eSBarry Smith     //ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr);
11915767789SHong Zhang     /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
12015767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
12115767789SHong Zhang   } else if (flag == MAT_GLOBAL_SUM) {
12215767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
12315767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
12415767789SHong Zhang     info->nz_used        = info->nz_allocated; /* assume Elemental does accurate allocation */
125ce94432eSBarry Smith     //ierr = MPI_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr);
12615767789SHong Zhang     //PetscPrintf(PETSC_COMM_SELF,"    ... [%d] locally allocated %g\n",rank,info->nz_allocated);
12715767789SHong Zhang   }
12815767789SHong Zhang 
12915767789SHong Zhang   info->nz_unneeded       = 0.0;
13015767789SHong Zhang   info->assemblies        = (double)A->num_ass;
13115767789SHong Zhang   info->mallocs           = 0;
13215767789SHong Zhang   info->memory            = ((PetscObject)A)->mem;
13315767789SHong Zhang   info->fill_ratio_given  = 0; /* determined by Elemental */
13415767789SHong Zhang   info->fill_ratio_needed = 0;
13515767789SHong Zhang   info->factor_mallocs    = 0;
136180a43e4SHong Zhang   PetscFunctionReturn(0);
137180a43e4SHong Zhang }
138180a43e4SHong Zhang 
139180a43e4SHong Zhang #undef __FUNCT__
140db31f6deSJed Brown #define __FUNCT__ "MatSetValues_Elemental"
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   PetscErrorCode ierr;
144db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
145db31f6deSJed Brown   PetscMPIInt    rank;
146db31f6deSJed Brown   PetscInt       i,j,rrank,ridx,crank,cidx;
147db31f6deSJed Brown 
148db31f6deSJed Brown   PetscFunctionBegin;
149ce94432eSBarry Smith   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr);
150db31f6deSJed Brown 
151db31f6deSJed Brown   const elem::Grid &grid = a->emat->Grid();
152db31f6deSJed Brown   for (i=0; i<nr; i++) {
153db31f6deSJed Brown     PetscInt erow,ecol,elrow,elcol;
154db31f6deSJed Brown     if (rows[i] < 0) continue;
155db31f6deSJed Brown     P2RO(A,0,rows[i],&rrank,&ridx);
156db31f6deSJed Brown     RO2E(A,0,rrank,ridx,&erow);
157ce94432eSBarry Smith     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
158db31f6deSJed Brown     for (j=0; j<nc; j++) {
159db31f6deSJed Brown       if (cols[j] < 0) continue;
160db31f6deSJed Brown       P2RO(A,1,cols[j],&crank,&cidx);
161db31f6deSJed Brown       RO2E(A,1,crank,cidx,&ecol);
162ce94432eSBarry Smith       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
163aae2c449SHong Zhang       if (erow % grid.MCSize() != grid.MCRank() || ecol % grid.MRSize() != grid.MRRank()){ /* off-proc entry */
164aae2c449SHong Zhang         if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
165aae2c449SHong Zhang         /* 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); */
166e6dea9dbSXuan Zhou         a->esubmat->Set(0,0, (PetscElemScalar)vals[i*nc+j]);
167aae2c449SHong Zhang         a->interface->Axpy(1.0,*(a->esubmat),erow,ecol);
168aae2c449SHong Zhang         continue;
169ed36708cSHong Zhang       }
170db31f6deSJed Brown       elrow = erow / grid.MCSize();
171db31f6deSJed Brown       elcol = ecol / grid.MRSize();
172db31f6deSJed Brown       switch (imode) {
173e6dea9dbSXuan Zhou       case INSERT_VALUES: a->emat->SetLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
174e6dea9dbSXuan Zhou       case ADD_VALUES: a->emat->UpdateLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
175ce94432eSBarry Smith       default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
176db31f6deSJed Brown       }
177db31f6deSJed Brown     }
178db31f6deSJed Brown   }
179db31f6deSJed Brown   PetscFunctionReturn(0);
180db31f6deSJed Brown }
181db31f6deSJed Brown 
182db31f6deSJed Brown #undef __FUNCT__
183db31f6deSJed Brown #define __FUNCT__ "MatMult_Elemental"
184db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
185db31f6deSJed Brown {
186db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
187db31f6deSJed Brown   PetscErrorCode        ierr;
188e6dea9dbSXuan Zhou   const PetscElemScalar *x;
189e6dea9dbSXuan Zhou   PetscElemScalar       *y;
190df311e6cSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
191db31f6deSJed Brown 
192db31f6deSJed Brown   PetscFunctionBegin;
193e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
194e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
195db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
196df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->cmap->N,1,0,x,A->cmap->n,*a->grid);
197df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ye(A->rmap->N,1,0,y,A->rmap->n,*a->grid);
198db31f6deSJed Brown     elem::Gemv(elem::NORMAL,one,*a->emat,xe,zero,ye);
199db31f6deSJed Brown   }
200e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
201e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
202db31f6deSJed Brown   PetscFunctionReturn(0);
203db31f6deSJed Brown }
204db31f6deSJed Brown 
205db31f6deSJed Brown #undef __FUNCT__
2069426833fSXuan Zhou #define __FUNCT__ "MatMultTranspose_Elemental"
2079426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
2089426833fSXuan Zhou {
2099426833fSXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
2109426833fSXuan Zhou   PetscErrorCode        ierr;
211df311e6cSXuan Zhou   const PetscElemScalar *x;
212df311e6cSXuan Zhou   PetscElemScalar       *y;
213e6dea9dbSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
2149426833fSXuan Zhou 
2159426833fSXuan Zhou   PetscFunctionBegin;
216e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
217e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2189426833fSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
219df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid);
220df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ye(A->cmap->N,1,0,y,A->cmap->n,*a->grid);
2219426833fSXuan Zhou     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,zero,ye);
2229426833fSXuan Zhou   }
223e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
224e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2259426833fSXuan Zhou   PetscFunctionReturn(0);
2269426833fSXuan Zhou }
2279426833fSXuan Zhou 
2289426833fSXuan Zhou #undef __FUNCT__
229db31f6deSJed Brown #define __FUNCT__ "MatMultAdd_Elemental"
230db31f6deSJed Brown static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
231db31f6deSJed Brown {
232db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
233db31f6deSJed Brown   PetscErrorCode        ierr;
234df311e6cSXuan Zhou   const PetscElemScalar *x;
235df311e6cSXuan Zhou   PetscElemScalar       *z;
236e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
237db31f6deSJed Brown 
238db31f6deSJed Brown   PetscFunctionBegin;
239db31f6deSJed Brown   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
240e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
241e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
242db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
243df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->cmap->N,1,0,x,A->cmap->n,*a->grid);
244df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ze(A->rmap->N,1,0,z,A->rmap->n,*a->grid);
245db31f6deSJed Brown     elem::Gemv(elem::NORMAL,one,*a->emat,xe,one,ze);
246db31f6deSJed Brown   }
247e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
248e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
249db31f6deSJed Brown   PetscFunctionReturn(0);
250db31f6deSJed Brown }
251db31f6deSJed Brown 
252db31f6deSJed Brown #undef __FUNCT__
253e883f9d5SXuan Zhou #define __FUNCT__ "MatMultTransposeAdd_Elemental"
254e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
255e883f9d5SXuan Zhou {
256e883f9d5SXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
257e883f9d5SXuan Zhou   PetscErrorCode        ierr;
258df311e6cSXuan Zhou   const PetscElemScalar *x;
259df311e6cSXuan Zhou   PetscElemScalar       *z;
260e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
261e883f9d5SXuan Zhou 
262e883f9d5SXuan Zhou   PetscFunctionBegin;
263e883f9d5SXuan Zhou   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
264e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
265e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
266e883f9d5SXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
267df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid);
268df311e6cSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ze(A->cmap->N,1,0,z,A->cmap->n,*a->grid);
269e883f9d5SXuan Zhou     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,one,ze);
270e883f9d5SXuan Zhou   }
271e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
272e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
273e883f9d5SXuan Zhou   PetscFunctionReturn(0);
274e883f9d5SXuan Zhou }
275e883f9d5SXuan Zhou 
276e883f9d5SXuan Zhou #undef __FUNCT__
2779a9e8502SHong Zhang #define __FUNCT__ "MatMatMultNumeric_Elemental"
2789a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
279c1d1b975SXuan Zhou {
280c1d1b975SXuan Zhou   Mat_Elemental    *a = (Mat_Elemental*)A->data;
281c1d1b975SXuan Zhou   Mat_Elemental    *b = (Mat_Elemental*)B->data;
2829a9e8502SHong Zhang   Mat_Elemental    *c = (Mat_Elemental*)C->data;
283e6dea9dbSXuan Zhou   PetscElemScalar  one = 1,zero = 0;
284c1d1b975SXuan Zhou 
285c1d1b975SXuan Zhou   PetscFunctionBegin;
286aae2c449SHong Zhang   { /* Scoping so that constructor is called before pointer is returned */
287c1d1b975SXuan Zhou     elem::Gemm(elem::NORMAL,elem::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
288aae2c449SHong Zhang   }
2899a9e8502SHong Zhang   C->assembled = PETSC_TRUE;
2909a9e8502SHong Zhang   PetscFunctionReturn(0);
2919a9e8502SHong Zhang }
2929a9e8502SHong Zhang 
2939a9e8502SHong Zhang #undef __FUNCT__
2949a9e8502SHong Zhang #define __FUNCT__ "MatMatMultSymbolic_Elemental"
2959a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
2969a9e8502SHong Zhang {
2979a9e8502SHong Zhang   PetscErrorCode ierr;
2989a9e8502SHong Zhang   Mat            Ce;
299ce94432eSBarry Smith   MPI_Comm       comm;
3009a9e8502SHong Zhang 
3019a9e8502SHong Zhang   PetscFunctionBegin;
302ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
3039a9e8502SHong Zhang   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
3049a9e8502SHong Zhang   ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
3059a9e8502SHong Zhang   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
3069a9e8502SHong Zhang   ierr = MatSetUp(Ce);CHKERRQ(ierr);
3079a9e8502SHong Zhang   *C = Ce;
3089a9e8502SHong Zhang   PetscFunctionReturn(0);
3099a9e8502SHong Zhang }
3109a9e8502SHong Zhang 
3119a9e8502SHong Zhang #undef __FUNCT__
3129a9e8502SHong Zhang #define __FUNCT__ "MatMatMult_Elemental"
3139a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3149a9e8502SHong Zhang {
3159a9e8502SHong Zhang   PetscErrorCode ierr;
3169a9e8502SHong Zhang 
3179a9e8502SHong Zhang   PetscFunctionBegin;
3189a9e8502SHong Zhang   if (scall == MAT_INITIAL_MATRIX){
3193ff4c91cSHong Zhang     ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3209a9e8502SHong Zhang     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
3213ff4c91cSHong Zhang     ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3229a9e8502SHong Zhang   }
3233ff4c91cSHong Zhang   ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
3249a9e8502SHong Zhang   ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
3253ff4c91cSHong Zhang   ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
326c1d1b975SXuan Zhou   PetscFunctionReturn(0);
327c1d1b975SXuan Zhou }
328c1d1b975SXuan Zhou 
329c1d1b975SXuan Zhou #undef __FUNCT__
330df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultNumeric_Elemental"
331df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
332df311e6cSXuan Zhou {
333df311e6cSXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
334df311e6cSXuan Zhou   Mat_Elemental      *b = (Mat_Elemental*)B->data;
335df311e6cSXuan Zhou   Mat_Elemental      *c = (Mat_Elemental*)C->data;
336e6dea9dbSXuan Zhou   PetscElemScalar    one = 1,zero = 0;
337df311e6cSXuan Zhou 
338df311e6cSXuan Zhou   PetscFunctionBegin;
339df311e6cSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
340df311e6cSXuan Zhou     elem::Gemm(elem::NORMAL,elem::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
341df311e6cSXuan Zhou   }
342df311e6cSXuan Zhou   C->assembled = PETSC_TRUE;
343df311e6cSXuan Zhou   PetscFunctionReturn(0);
344df311e6cSXuan Zhou }
345df311e6cSXuan Zhou 
346df311e6cSXuan Zhou #undef __FUNCT__
347df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultSymbolic_Elemental"
348df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
349df311e6cSXuan Zhou {
350df311e6cSXuan Zhou   PetscErrorCode ierr;
351df311e6cSXuan Zhou   Mat            Ce;
352ce94432eSBarry Smith   MPI_Comm       comm;
353df311e6cSXuan Zhou 
354df311e6cSXuan Zhou   PetscFunctionBegin;
355ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
356df311e6cSXuan Zhou   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
357df311e6cSXuan Zhou   ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
358df311e6cSXuan Zhou   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
359df311e6cSXuan Zhou   ierr = MatSetUp(Ce);CHKERRQ(ierr);
360df311e6cSXuan Zhou   *C = Ce;
361df311e6cSXuan Zhou   PetscFunctionReturn(0);
362df311e6cSXuan Zhou }
363df311e6cSXuan Zhou 
364df311e6cSXuan Zhou #undef __FUNCT__
365df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMult_Elemental"
366df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
367df311e6cSXuan Zhou {
368df311e6cSXuan Zhou   PetscErrorCode ierr;
369df311e6cSXuan Zhou 
370df311e6cSXuan Zhou   PetscFunctionBegin;
371df311e6cSXuan Zhou   if (scall == MAT_INITIAL_MATRIX){
372df311e6cSXuan Zhou     ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
373df311e6cSXuan Zhou     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
374df311e6cSXuan Zhou     ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
375df311e6cSXuan Zhou   }
376df311e6cSXuan Zhou   ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
377df311e6cSXuan Zhou   ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
378df311e6cSXuan Zhou   ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
379df311e6cSXuan Zhou   PetscFunctionReturn(0);
380df311e6cSXuan Zhou }
381df311e6cSXuan Zhou 
382df311e6cSXuan Zhou #undef __FUNCT__
38361119200SXuan Zhou #define __FUNCT__ "MatGetDiagonal_Elemental"
384a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
38561119200SXuan Zhou {
386a9d89745SXuan Zhou   PetscInt        i,nrows,ncols,nD,rrank,ridx,crank,cidx;
387a9d89745SXuan Zhou   Mat_Elemental   *a = (Mat_Elemental*)A->data;
38861119200SXuan Zhou   PetscErrorCode  ierr;
389a9d89745SXuan Zhou   PetscElemScalar v;
390ce94432eSBarry Smith   MPI_Comm        comm;
39161119200SXuan Zhou 
39261119200SXuan Zhou   PetscFunctionBegin;
393ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
394a9d89745SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
395a9d89745SXuan Zhou   nD = nrows>ncols ? ncols : nrows;
396a9d89745SXuan Zhou   for (i=0; i<nD; i++) {
397a9d89745SXuan Zhou     PetscInt erow,ecol;
398a9d89745SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
399a9d89745SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
400a9d89745SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
401a9d89745SXuan Zhou     P2RO(A,1,i,&crank,&cidx);
402a9d89745SXuan Zhou     RO2E(A,1,crank,cidx,&ecol);
403a9d89745SXuan Zhou     if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
404a9d89745SXuan Zhou     v = a->emat->Get(erow,ecol);
4057c8b904fSSatish Balay     ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr);
406ade3cc5eSXuan Zhou   }
407a9d89745SXuan Zhou   ierr = VecAssemblyBegin(D);CHKERRQ(ierr);
408a9d89745SXuan Zhou   ierr = VecAssemblyEnd(D);CHKERRQ(ierr);
40961119200SXuan Zhou   PetscFunctionReturn(0);
41061119200SXuan Zhou }
41161119200SXuan Zhou 
41261119200SXuan Zhou #undef __FUNCT__
413ade3cc5eSXuan Zhou #define __FUNCT__ "MatDiagonalScale_Elemental"
414ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
415ade3cc5eSXuan Zhou {
416ade3cc5eSXuan Zhou   Mat_Elemental         *x = (Mat_Elemental*)X->data;
417ade3cc5eSXuan Zhou   const PetscElemScalar *d;
418ade3cc5eSXuan Zhou   PetscErrorCode        ierr;
419ade3cc5eSXuan Zhou 
420ade3cc5eSXuan Zhou   PetscFunctionBegin;
4219065cd98SJed Brown   if (R) {
422ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
423ade3cc5eSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de(X->cmap->N,1,0,d,X->cmap->n,*x->grid);
424ade3cc5eSXuan Zhou     elem::DiagonalScale(elem::RIGHT,elem::NORMAL,de,*x->emat);
425ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
4269065cd98SJed Brown   }
4279065cd98SJed Brown   if (L) {
428ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
429ade3cc5eSXuan Zhou     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de(X->rmap->N,1,0,d,X->rmap->n,*x->grid);
430ade3cc5eSXuan Zhou     elem::DiagonalScale(elem::LEFT,elem::NORMAL,de,*x->emat);
431ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
432ade3cc5eSXuan Zhou   }
433ade3cc5eSXuan Zhou   PetscFunctionReturn(0);
434ade3cc5eSXuan Zhou }
435ade3cc5eSXuan Zhou 
436ade3cc5eSXuan Zhou #undef __FUNCT__
4374ee44ac6SXuan Zhou #define __FUNCT__ "MatScale_Elemental"
438e6dea9dbSXuan Zhou static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a)
43965b78793SXuan Zhou {
44065b78793SXuan Zhou   Mat_Elemental  *x = (Mat_Elemental*)X->data;
44165b78793SXuan Zhou 
44265b78793SXuan Zhou   PetscFunctionBegin;
443e6dea9dbSXuan Zhou   elem::Scal((PetscElemScalar)a,*x->emat);
44465b78793SXuan Zhou   PetscFunctionReturn(0);
44565b78793SXuan Zhou }
44665b78793SXuan Zhou 
44765b78793SXuan Zhou #undef __FUNCT__
4484ee44ac6SXuan Zhou #define __FUNCT__ "MatAXPY_Elemental"
449e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
450e09a3074SHong Zhang {
451e09a3074SHong Zhang   Mat_Elemental  *x = (Mat_Elemental*)X->data;
452e09a3074SHong Zhang   Mat_Elemental  *y = (Mat_Elemental*)Y->data;
453e09a3074SHong Zhang 
454e09a3074SHong Zhang   PetscFunctionBegin;
455e6dea9dbSXuan Zhou   elem::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
456e09a3074SHong Zhang   PetscFunctionReturn(0);
457e09a3074SHong Zhang }
458e09a3074SHong Zhang 
459ae844d54SHong Zhang #undef __FUNCT__
460d6223691SXuan Zhou #define __FUNCT__ "MatCopy_Elemental"
461d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
462d6223691SXuan Zhou {
463d6223691SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
464d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
465d6223691SXuan Zhou 
466d6223691SXuan Zhou   PetscFunctionBegin;
467d6223691SXuan Zhou   elem::Copy(*a->emat,*b->emat);
468d6223691SXuan Zhou   PetscFunctionReturn(0);
469d6223691SXuan Zhou }
470d6223691SXuan Zhou 
471d6223691SXuan Zhou #undef __FUNCT__
472df311e6cSXuan Zhou #define __FUNCT__ "MatDuplicate_Elemental"
473df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
474df311e6cSXuan Zhou {
475df311e6cSXuan Zhou   Mat            Be;
476ce94432eSBarry Smith   MPI_Comm       comm;
477df311e6cSXuan Zhou   Mat_Elemental  *a=(Mat_Elemental*)A->data;
478df311e6cSXuan Zhou   PetscErrorCode ierr;
479df311e6cSXuan Zhou 
480df311e6cSXuan Zhou   PetscFunctionBegin;
481ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
482df311e6cSXuan Zhou   ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
483df311e6cSXuan Zhou   ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
484df311e6cSXuan Zhou   ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
485df311e6cSXuan Zhou   ierr = MatSetUp(Be);CHKERRQ(ierr);
486df311e6cSXuan Zhou   *B = Be;
487df311e6cSXuan Zhou   if (op == MAT_COPY_VALUES) {
488df311e6cSXuan Zhou     Mat_Elemental *b=(Mat_Elemental*)Be->data;
489df311e6cSXuan Zhou     elem::Copy(*a->emat,*b->emat);
490df311e6cSXuan Zhou   }
491df311e6cSXuan Zhou   Be->assembled = PETSC_TRUE;
492df311e6cSXuan Zhou   PetscFunctionReturn(0);
493df311e6cSXuan Zhou }
494df311e6cSXuan Zhou 
495df311e6cSXuan Zhou #undef __FUNCT__
496d6223691SXuan Zhou #define __FUNCT__ "MatTranspose_Elemental"
497d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
498d6223691SXuan Zhou {
4995262d616SXuan Zhou   Mat            Be;
5005262d616SXuan Zhou   PetscErrorCode ierr;
501ce94432eSBarry Smith   MPI_Comm       comm;
5024fe7bbcaSHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
503d6223691SXuan Zhou 
504d6223691SXuan Zhou   PetscFunctionBegin;
505ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5065cb544a0SHong Zhang   /* Only out-of-place supported */
5075262d616SXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5085262d616SXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5095262d616SXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5105262d616SXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5115262d616SXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5125262d616SXuan Zhou     *B = Be;
5135262d616SXuan Zhou   }
5144fe7bbcaSHong Zhang   b = (Mat_Elemental*)Be->data;
5155262d616SXuan Zhou   elem::Transpose(*a->emat,*b->emat);
5165262d616SXuan Zhou   Be->assembled = PETSC_TRUE;
517d6223691SXuan Zhou   PetscFunctionReturn(0);
518d6223691SXuan Zhou }
519d6223691SXuan Zhou 
520d6223691SXuan Zhou #undef __FUNCT__
521dfcb0403SXuan Zhou #define __FUNCT__ "MatConjugate_Elemental"
522dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A)
523dfcb0403SXuan Zhou {
524dfcb0403SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
525dfcb0403SXuan Zhou 
526dfcb0403SXuan Zhou   PetscFunctionBegin;
527dfcb0403SXuan Zhou   elem::Conjugate(*a->emat);
528dfcb0403SXuan Zhou   PetscFunctionReturn(0);
529dfcb0403SXuan Zhou }
530dfcb0403SXuan Zhou 
531dfcb0403SXuan Zhou #undef __FUNCT__
5324a29722dSXuan Zhou #define __FUNCT__ "MatHermitianTranspose_Elemental"
5334a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
5344a29722dSXuan Zhou {
5354a29722dSXuan Zhou   Mat            Be;
5364a29722dSXuan Zhou   PetscErrorCode ierr;
537ce94432eSBarry Smith   MPI_Comm       comm;
5384a29722dSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
5394a29722dSXuan Zhou 
5404a29722dSXuan Zhou   PetscFunctionBegin;
541ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5425cb544a0SHong Zhang   /* Only out-of-place supported */
5434a29722dSXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5444a29722dSXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5454a29722dSXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5464a29722dSXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5474a29722dSXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5484a29722dSXuan Zhou     *B = Be;
5494a29722dSXuan Zhou   }
5504a29722dSXuan Zhou   b = (Mat_Elemental*)Be->data;
5514a29722dSXuan Zhou   elem::Adjoint(*a->emat,*b->emat);
5524a29722dSXuan Zhou   Be->assembled = PETSC_TRUE;
5534a29722dSXuan Zhou   PetscFunctionReturn(0);
5544a29722dSXuan Zhou }
5554a29722dSXuan Zhou 
5564a29722dSXuan Zhou #undef __FUNCT__
5571f881ff8SXuan Zhou #define __FUNCT__ "MatSolve_Elemental"
5581f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
5591f881ff8SXuan Zhou {
5601f881ff8SXuan Zhou   Mat_Elemental     *a = (Mat_Elemental*)A->data;
5611f881ff8SXuan Zhou   PetscErrorCode    ierr;
562df311e6cSXuan Zhou   PetscElemScalar   *x;
5631f881ff8SXuan Zhou 
5641f881ff8SXuan Zhou   PetscFunctionBegin;
56545cf121fSXuan Zhou   ierr = VecCopy(B,X);CHKERRQ(ierr);
566e6dea9dbSXuan Zhou   ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
567df311e6cSXuan Zhou   elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid);
568df311e6cSXuan Zhou   elem::DistMatrix<PetscElemScalar,elem::MC,elem::MR> xer = xe;
569fc54b460SXuan Zhou   switch (A->factortype) {
570fc54b460SXuan Zhou   case MAT_FACTOR_LU:
5711f881ff8SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
572324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,xer);
57345cf121fSXuan Zhou       elem::Copy(xer,xe);
574fc54b460SXuan Zhou     } else {
575324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,xer);
57645cf121fSXuan Zhou       elem::Copy(xer,xe);
5771f881ff8SXuan Zhou     }
578fc54b460SXuan Zhou     break;
579fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
580324645c7SJack Poulson     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,xer);
581fc54b460SXuan Zhou     elem::Copy(xer,xe);
582fc54b460SXuan Zhou     break;
583fc54b460SXuan Zhou   default:
5844fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
585fc54b460SXuan Zhou     break;
5861f881ff8SXuan Zhou   }
587e6dea9dbSXuan Zhou   ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
5881f881ff8SXuan Zhou   PetscFunctionReturn(0);
5891f881ff8SXuan Zhou }
5901f881ff8SXuan Zhou 
5911f881ff8SXuan Zhou #undef __FUNCT__
592df311e6cSXuan Zhou #define __FUNCT__ "MatSolveAdd_Elemental"
593df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
594df311e6cSXuan Zhou {
595df311e6cSXuan Zhou   PetscErrorCode    ierr;
596df311e6cSXuan Zhou 
597df311e6cSXuan Zhou   PetscFunctionBegin;
598df311e6cSXuan Zhou   ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr);
5993d7f40dbSXuan Zhou   ierr = VecAXPY(X,1,Y);CHKERRQ(ierr);
600df311e6cSXuan Zhou   PetscFunctionReturn(0);
601df311e6cSXuan Zhou }
602df311e6cSXuan Zhou 
603df311e6cSXuan Zhou #undef __FUNCT__
604ae844d54SHong Zhang #define __FUNCT__ "MatMatSolve_Elemental"
605ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
606ae844d54SHong Zhang {
6071f0e42cfSHong Zhang   Mat_Elemental *a=(Mat_Elemental*)A->data;
608d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
6091f0e42cfSHong Zhang   Mat_Elemental *x=(Mat_Elemental*)X->data;
6101f0e42cfSHong Zhang 
611ae844d54SHong Zhang   PetscFunctionBegin;
612d6223691SXuan Zhou   elem::Copy(*b->emat,*x->emat);
613fc54b460SXuan Zhou   switch (A->factortype) {
614fc54b460SXuan Zhou   case MAT_FACTOR_LU:
615d6223691SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
616324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,*x->emat);
617fc54b460SXuan Zhou     } else {
618324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*x->emat);
619d6223691SXuan Zhou     }
620fc54b460SXuan Zhou     break;
621fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
622324645c7SJack Poulson     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,*x->emat);
623fc54b460SXuan Zhou     break;
624fc54b460SXuan Zhou   default:
6254fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
626fc54b460SXuan Zhou     break;
627fc54b460SXuan Zhou   }
628ae844d54SHong Zhang   PetscFunctionReturn(0);
629ae844d54SHong Zhang }
630ae844d54SHong Zhang 
631ae844d54SHong Zhang #undef __FUNCT__
632ae844d54SHong Zhang #define __FUNCT__ "MatLUFactor_Elemental"
633ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
634ae844d54SHong Zhang {
6357c920d81SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
6367c920d81SXuan Zhou 
637ae844d54SHong Zhang   PetscFunctionBegin;
638d6223691SXuan Zhou   if (info->dtcol){
6397c920d81SXuan Zhou     elem::LU(*a->emat,*a->pivot);
6401293973dSHong Zhang   } else {
641d6223691SXuan Zhou     elem::LU(*a->emat);
642d6223691SXuan Zhou   }
6431293973dSHong Zhang   A->factortype = MAT_FACTOR_LU;
644834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
645ae844d54SHong Zhang   PetscFunctionReturn(0);
646ae844d54SHong Zhang }
647ae844d54SHong Zhang 
648d7c3f9d8SHong Zhang #undef __FUNCT__
649d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorNumeric_Elemental"
650d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
651d7c3f9d8SHong Zhang {
652d7c3f9d8SHong Zhang   PetscErrorCode ierr;
653d7c3f9d8SHong Zhang 
654d7c3f9d8SHong Zhang   PetscFunctionBegin;
655d7c3f9d8SHong Zhang   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
656d7c3f9d8SHong Zhang   ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr);
657d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
658d7c3f9d8SHong Zhang }
659d7c3f9d8SHong Zhang 
660d7c3f9d8SHong Zhang #undef __FUNCT__
661d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorSymbolic_Elemental"
662d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
663d7c3f9d8SHong Zhang {
664d7c3f9d8SHong Zhang   PetscFunctionBegin;
665d7c3f9d8SHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
666d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
667d7c3f9d8SHong Zhang }
668d7c3f9d8SHong Zhang 
66945cf121fSXuan Zhou #undef __FUNCT__
67045cf121fSXuan Zhou #define __FUNCT__ "MatCholeskyFactor_Elemental"
67145cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
67245cf121fSXuan Zhou {
67345cf121fSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
674df311e6cSXuan Zhou   elem::DistMatrix<PetscElemScalar,elem::MC,elem::STAR> d;
67545cf121fSXuan Zhou 
67645cf121fSXuan Zhou   PetscFunctionBegin;
677c9fc186eSXuan Zhou   elem::Cholesky(elem::UPPER,*a->emat);
678fc54b460SXuan Zhou   A->factortype = MAT_FACTOR_CHOLESKY;
679834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
68045cf121fSXuan Zhou   PetscFunctionReturn(0);
68145cf121fSXuan Zhou }
68245cf121fSXuan Zhou 
68379673f7bSHong Zhang #undef __FUNCT__
68479673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorNumeric_Elemental"
68579673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
68679673f7bSHong Zhang {
687cb76c1d8SXuan Zhou   PetscErrorCode ierr;
688cb76c1d8SXuan Zhou 
689cb76c1d8SXuan Zhou   PetscFunctionBegin;
690cb76c1d8SXuan Zhou   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
691cb76c1d8SXuan Zhou   ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr);
692cb76c1d8SXuan Zhou   PetscFunctionReturn(0);
69379673f7bSHong Zhang }
694cb76c1d8SXuan Zhou 
69579673f7bSHong Zhang #undef __FUNCT__
69679673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorSymbolic_Elemental"
69779673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
69879673f7bSHong Zhang {
69979673f7bSHong Zhang   PetscFunctionBegin;
70079673f7bSHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
70179673f7bSHong Zhang   PetscFunctionReturn(0);
70279673f7bSHong Zhang }
70379673f7bSHong Zhang 
7041293973dSHong Zhang #undef __FUNCT__
70515767789SHong Zhang #define __FUNCT__ "MatFactorGetSolverPackage_elemental_elemental"
70615767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type)
70715767789SHong Zhang {
70815767789SHong Zhang   PetscFunctionBegin;
70915767789SHong Zhang   *type = MATSOLVERELEMENTAL;
71015767789SHong Zhang   PetscFunctionReturn(0);
71115767789SHong Zhang }
71215767789SHong Zhang 
71315767789SHong Zhang #undef __FUNCT__
71415767789SHong Zhang #define __FUNCT__ "MatGetFactor_elemental_elemental"
71515767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
7161293973dSHong Zhang {
7171293973dSHong Zhang   Mat            B;
7181293973dSHong Zhang   PetscErrorCode ierr;
7191293973dSHong Zhang 
7201293973dSHong Zhang   PetscFunctionBegin;
7211293973dSHong Zhang   /* Create the factorization matrix */
722ce94432eSBarry Smith   ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
7231293973dSHong Zhang   ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
7241293973dSHong Zhang   ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr);
7251293973dSHong Zhang   ierr = MatSetUp(B);CHKERRQ(ierr);
7261293973dSHong Zhang   B->factortype = ftype;
727bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr);
7281293973dSHong Zhang   *F            = B;
7291293973dSHong Zhang   PetscFunctionReturn(0);
7301293973dSHong Zhang }
7311293973dSHong Zhang 
7321f881ff8SXuan Zhou #undef __FUNCT__
7331f881ff8SXuan Zhou #define __FUNCT__ "MatNorm_Elemental"
7341f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
7351f881ff8SXuan Zhou {
7361f881ff8SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7371f881ff8SXuan Zhou 
7381f881ff8SXuan Zhou   PetscFunctionBegin;
7391f881ff8SXuan Zhou   switch (type){
7401f881ff8SXuan Zhou   case NORM_1:
741324645c7SJack Poulson     *nrm = elem::OneNorm(*a->emat);
7421f881ff8SXuan Zhou     break;
7431f881ff8SXuan Zhou   case NORM_FROBENIUS:
744324645c7SJack Poulson     *nrm = elem::FrobeniusNorm(*a->emat);
7451f881ff8SXuan Zhou     break;
7461f881ff8SXuan Zhou   case NORM_INFINITY:
747324645c7SJack Poulson     *nrm = elem::InfinityNorm(*a->emat);
7481f881ff8SXuan Zhou     break;
7491f881ff8SXuan Zhou   default:
7501f881ff8SXuan Zhou     printf("Error: unsupported norm type!\n");
7511f881ff8SXuan Zhou   }
7521f881ff8SXuan Zhou   PetscFunctionReturn(0);
7531f881ff8SXuan Zhou }
7541f881ff8SXuan Zhou 
7555262d616SXuan Zhou #undef __FUNCT__
7565262d616SXuan Zhou #define __FUNCT__ "MatZeroEntries_Elemental"
7575262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A)
7585262d616SXuan Zhou {
7595262d616SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7605262d616SXuan Zhou 
7615262d616SXuan Zhou   PetscFunctionBegin;
7625262d616SXuan Zhou   elem::Zero(*a->emat);
7635262d616SXuan Zhou   PetscFunctionReturn(0);
7645262d616SXuan Zhou }
7655262d616SXuan Zhou 
766e09a3074SHong Zhang #undef __FUNCT__
767db31f6deSJed Brown #define __FUNCT__ "MatGetOwnershipIS_Elemental"
768db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
769db31f6deSJed Brown {
770db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
771db31f6deSJed Brown   PetscErrorCode ierr;
772db31f6deSJed Brown   PetscInt       i,m,shift,stride,*idx;
773db31f6deSJed Brown 
774db31f6deSJed Brown   PetscFunctionBegin;
775db31f6deSJed Brown   if (rows) {
776db31f6deSJed Brown     m = a->emat->LocalHeight();
777db31f6deSJed Brown     shift = a->emat->ColShift();
778db31f6deSJed Brown     stride = a->emat->ColStride();
779785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
780db31f6deSJed Brown     for (i=0; i<m; i++) {
781db31f6deSJed Brown       PetscInt rank,offset;
782db31f6deSJed Brown       E2RO(A,0,shift+i*stride,&rank,&offset);
783db31f6deSJed Brown       RO2P(A,0,rank,offset,&idx[i]);
784db31f6deSJed Brown     }
785db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
786db31f6deSJed Brown   }
787db31f6deSJed Brown   if (cols) {
788db31f6deSJed Brown     m = a->emat->LocalWidth();
789db31f6deSJed Brown     shift = a->emat->RowShift();
790db31f6deSJed Brown     stride = a->emat->RowStride();
791785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
792db31f6deSJed Brown     for (i=0; i<m; i++) {
793db31f6deSJed Brown       PetscInt rank,offset;
794db31f6deSJed Brown       E2RO(A,1,shift+i*stride,&rank,&offset);
795db31f6deSJed Brown       RO2P(A,1,rank,offset,&idx[i]);
796db31f6deSJed Brown     }
797db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr);
798db31f6deSJed Brown   }
799db31f6deSJed Brown   PetscFunctionReturn(0);
800db31f6deSJed Brown }
801db31f6deSJed Brown 
802db31f6deSJed Brown #undef __FUNCT__
8032ef0cf24SXuan Zhou #define __FUNCT__ "MatConvert_Elemental_Dense"
80419fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
805af295397SXuan Zhou {
8062ef0cf24SXuan Zhou   Mat                Bmpi;
807af295397SXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
808ce94432eSBarry Smith   MPI_Comm           comm;
8092ef0cf24SXuan Zhou   PetscErrorCode     ierr;
8102ef0cf24SXuan Zhou   PetscInt           rrank,ridx,crank,cidx,nrows,ncols,i,j;
811df311e6cSXuan Zhou   PetscElemScalar    v;
812573b0fb4SBarry Smith   PetscBool          s1,s2,s3;
813af295397SXuan Zhou 
814af295397SXuan Zhou   PetscFunctionBegin;
815ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
816573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATDENSE,&s1);CHKERRQ(ierr);
817573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATSEQDENSE,&s2);CHKERRQ(ierr);
818573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATMPIDENSE,&s3);CHKERRQ(ierr);
819573b0fb4SBarry Smith   if (!s1 && !s2 && !s3) SETERRQ(comm,PETSC_ERR_SUP,"Unsupported New MatType: must be MATDENSE, MATSEQDENSE or MATMPIDENSE");
820af295397SXuan Zhou   ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr);
821af295397SXuan Zhou   ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
8222ef0cf24SXuan Zhou   ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr);
823af295397SXuan Zhou   ierr = MatSetUp(Bmpi);CHKERRQ(ierr);
8242ef0cf24SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
8252ef0cf24SXuan Zhou   for (i=0; i<nrows; i++) {
8262ef0cf24SXuan Zhou     PetscInt erow,ecol;
8272ef0cf24SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
8282ef0cf24SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
8292ef0cf24SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
8302ef0cf24SXuan Zhou     for (j=0; j<ncols; j++) {
8312ef0cf24SXuan Zhou       P2RO(A,1,j,&crank,&cidx);
8322ef0cf24SXuan Zhou       RO2E(A,1,crank,cidx,&ecol);
8332ef0cf24SXuan Zhou       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
8342ef0cf24SXuan Zhou       v = a->emat->Get(erow,ecol);
835e6dea9dbSXuan Zhou       ierr = MatSetValues(Bmpi,1,&i,1,&j,(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
8362ef0cf24SXuan Zhou     }
8372ef0cf24SXuan Zhou   }
838af295397SXuan Zhou   ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
839af295397SXuan Zhou   ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
840c4ad791aSXuan Zhou   if (reuse == MAT_REUSE_MATRIX) {
841c4ad791aSXuan Zhou     ierr = MatHeaderReplace(A,Bmpi);CHKERRQ(ierr);
842c4ad791aSXuan Zhou   } else {
843c4ad791aSXuan Zhou     *B = Bmpi;
844c4ad791aSXuan Zhou   }
845af295397SXuan Zhou   PetscFunctionReturn(0);
846af295397SXuan Zhou }
847af295397SXuan Zhou 
848af295397SXuan Zhou #undef __FUNCT__
849db31f6deSJed Brown #define __FUNCT__ "MatDestroy_Elemental"
850db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A)
851db31f6deSJed Brown {
852db31f6deSJed Brown   Mat_Elemental      *a = (Mat_Elemental*)A->data;
853db31f6deSJed Brown   PetscErrorCode     ierr;
8545e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
8555e9f5b67SHong Zhang   PetscBool          flg;
8565e9f5b67SHong Zhang   MPI_Comm           icomm;
857db31f6deSJed Brown 
858db31f6deSJed Brown   PetscFunctionBegin;
859c1ee1e62SHong Zhang   a->interface->Detach();
860aae2c449SHong Zhang   delete a->interface;
861aae2c449SHong Zhang   delete a->esubmat;
862db31f6deSJed Brown   delete a->emat;
8635e9f5b67SHong Zhang 
864ce94432eSBarry Smith   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
8650298fd71SBarry Smith   ierr = PetscCommDuplicate(cxxcomm,&icomm,NULL);CHKERRQ(ierr);
8665e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
8675e9f5b67SHong Zhang   if (--commgrid->grid_refct == 0) {
8685e9f5b67SHong Zhang     delete commgrid->grid;
8695e9f5b67SHong Zhang     ierr = PetscFree(commgrid);CHKERRQ(ierr);
8705e9f5b67SHong Zhang   }
8715e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
872bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr);
873bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetFactor_petsc_C",NULL);CHKERRQ(ierr);
874bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr);
875db31f6deSJed Brown   ierr = PetscFree(A->data);CHKERRQ(ierr);
876db31f6deSJed Brown   PetscFunctionReturn(0);
877db31f6deSJed Brown }
878db31f6deSJed Brown 
879db31f6deSJed Brown #undef __FUNCT__
880db31f6deSJed Brown #define __FUNCT__ "MatSetUp_Elemental"
881db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A)
882db31f6deSJed Brown {
883db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
884db31f6deSJed Brown   PetscErrorCode ierr;
885db31f6deSJed Brown   PetscMPIInt    rsize,csize;
886db31f6deSJed Brown 
887db31f6deSJed Brown   PetscFunctionBegin;
888db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
889db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
890db31f6deSJed Brown 
891db31f6deSJed Brown   a->emat->ResizeTo(A->rmap->N,A->cmap->N);CHKERRQ(ierr);
892db31f6deSJed Brown   elem::Zero(*a->emat);
893db31f6deSJed Brown 
894db31f6deSJed Brown   ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr);
895db31f6deSJed Brown   ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr);
896ce94432eSBarry Smith   if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
897db31f6deSJed Brown   a->commsize = rsize;
898db31f6deSJed Brown   a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
899db31f6deSJed Brown   a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
900db31f6deSJed Brown   a->m[0]  = A->rmap->N / rsize + (a->mr[0] != rsize);
901db31f6deSJed Brown   a->m[1]  = A->cmap->N / csize + (a->mr[1] != csize);
902db31f6deSJed Brown   PetscFunctionReturn(0);
903db31f6deSJed Brown }
904db31f6deSJed Brown 
905aae2c449SHong Zhang #undef __FUNCT__
906aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyBegin_Elemental"
907aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
908aae2c449SHong Zhang {
909aae2c449SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
910aae2c449SHong Zhang 
911aae2c449SHong Zhang   PetscFunctionBegin;
912aae2c449SHong Zhang   a->interface->Detach();
913aae2c449SHong Zhang   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
914aae2c449SHong Zhang   PetscFunctionReturn(0);
915aae2c449SHong Zhang }
916aae2c449SHong Zhang 
917aae2c449SHong Zhang #undef __FUNCT__
918aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyEnd_Elemental"
919aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
920aae2c449SHong Zhang {
921aae2c449SHong Zhang   PetscFunctionBegin;
922aae2c449SHong Zhang   /* Currently does nothing */
923aae2c449SHong Zhang   PetscFunctionReturn(0);
924aae2c449SHong Zhang }
925aae2c449SHong Zhang 
92640d92e34SHong Zhang /* -------------------------------------------------------------------*/
92740d92e34SHong Zhang static struct _MatOps MatOps_Values = {
92840d92e34SHong Zhang        MatSetValues_Elemental,
92940d92e34SHong Zhang        0,
93040d92e34SHong Zhang        0,
93140d92e34SHong Zhang        MatMult_Elemental,
93240d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental,
9339426833fSXuan Zhou        MatMultTranspose_Elemental,
934e883f9d5SXuan Zhou        MatMultTransposeAdd_Elemental,
93540d92e34SHong Zhang        MatSolve_Elemental,
936df311e6cSXuan Zhou        MatSolveAdd_Elemental,
93740d92e34SHong Zhang        0, //MatSolveTranspose_Elemental,
93840d92e34SHong Zhang /*10*/ 0, //MatSolveTransposeAdd_Elemental,
93940d92e34SHong Zhang        MatLUFactor_Elemental,
94040d92e34SHong Zhang        MatCholeskyFactor_Elemental,
94140d92e34SHong Zhang        0,
94240d92e34SHong Zhang        MatTranspose_Elemental,
94340d92e34SHong Zhang /*15*/ MatGetInfo_Elemental,
94440d92e34SHong Zhang        0,
94561119200SXuan Zhou        MatGetDiagonal_Elemental,
946ade3cc5eSXuan Zhou        MatDiagonalScale_Elemental,
94740d92e34SHong Zhang        MatNorm_Elemental,
94840d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental,
94940d92e34SHong Zhang        MatAssemblyEnd_Elemental,
95040d92e34SHong Zhang        0, //MatSetOption_Elemental,
95140d92e34SHong Zhang        MatZeroEntries_Elemental,
95240d92e34SHong Zhang /*24*/ 0,
95340d92e34SHong Zhang        MatLUFactorSymbolic_Elemental,
95440d92e34SHong Zhang        MatLUFactorNumeric_Elemental,
95540d92e34SHong Zhang        MatCholeskyFactorSymbolic_Elemental,
95640d92e34SHong Zhang        MatCholeskyFactorNumeric_Elemental,
95740d92e34SHong Zhang /*29*/ MatSetUp_Elemental,
95840d92e34SHong Zhang        0,
95940d92e34SHong Zhang        0,
96040d92e34SHong Zhang        0,
96140d92e34SHong Zhang        0,
962df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental,
96340d92e34SHong Zhang        0,
96440d92e34SHong Zhang        0,
96540d92e34SHong Zhang        0,
96640d92e34SHong Zhang        0,
96740d92e34SHong Zhang /*39*/ MatAXPY_Elemental,
96840d92e34SHong Zhang        0,
96940d92e34SHong Zhang        0,
97040d92e34SHong Zhang        0,
97140d92e34SHong Zhang        MatCopy_Elemental,
97240d92e34SHong Zhang /*44*/ 0,
97340d92e34SHong Zhang        MatScale_Elemental,
97440d92e34SHong Zhang        0,
97540d92e34SHong Zhang        0,
97640d92e34SHong Zhang        0,
97740d92e34SHong Zhang /*49*/ 0,
97840d92e34SHong Zhang        0,
97940d92e34SHong Zhang        0,
98040d92e34SHong Zhang        0,
98140d92e34SHong Zhang        0,
98240d92e34SHong Zhang /*54*/ 0,
98340d92e34SHong Zhang        0,
98440d92e34SHong Zhang        0,
98540d92e34SHong Zhang        0,
98640d92e34SHong Zhang        0,
98740d92e34SHong Zhang /*59*/ 0,
98840d92e34SHong Zhang        MatDestroy_Elemental,
98940d92e34SHong Zhang        MatView_Elemental,
99040d92e34SHong Zhang        0,
99140d92e34SHong Zhang        0,
99240d92e34SHong Zhang /*64*/ 0,
99340d92e34SHong Zhang        0,
99440d92e34SHong Zhang        0,
99540d92e34SHong Zhang        0,
99640d92e34SHong Zhang        0,
99740d92e34SHong Zhang /*69*/ 0,
99840d92e34SHong Zhang        0,
9992ef0cf24SXuan Zhou        MatConvert_Elemental_Dense,
100040d92e34SHong Zhang        0,
100140d92e34SHong Zhang        0,
100240d92e34SHong Zhang /*74*/ 0,
100340d92e34SHong Zhang        0,
100440d92e34SHong Zhang        0,
100540d92e34SHong Zhang        0,
100640d92e34SHong Zhang        0,
100740d92e34SHong Zhang /*79*/ 0,
100840d92e34SHong Zhang        0,
100940d92e34SHong Zhang        0,
101040d92e34SHong Zhang        0,
101140d92e34SHong Zhang        0,
101240d92e34SHong Zhang /*84*/ 0,
101340d92e34SHong Zhang        0,
101440d92e34SHong Zhang        0,
101540d92e34SHong Zhang        0,
101640d92e34SHong Zhang        0,
101740d92e34SHong Zhang /*89*/ MatMatMult_Elemental,
101840d92e34SHong Zhang        MatMatMultSymbolic_Elemental,
101940d92e34SHong Zhang        MatMatMultNumeric_Elemental,
102040d92e34SHong Zhang        0,
102140d92e34SHong Zhang        0,
102240d92e34SHong Zhang /*94*/ 0,
1023df311e6cSXuan Zhou        MatMatTransposeMult_Elemental,
1024df311e6cSXuan Zhou        MatMatTransposeMultSymbolic_Elemental,
1025df311e6cSXuan Zhou        MatMatTransposeMultNumeric_Elemental,
102640d92e34SHong Zhang        0,
102740d92e34SHong Zhang /*99*/ 0,
102840d92e34SHong Zhang        0,
102940d92e34SHong Zhang        0,
1030dfcb0403SXuan Zhou        MatConjugate_Elemental,
103140d92e34SHong Zhang        0,
103240d92e34SHong Zhang /*104*/0,
103340d92e34SHong Zhang        0,
103440d92e34SHong Zhang        0,
103540d92e34SHong Zhang        0,
103640d92e34SHong Zhang        0,
103740d92e34SHong Zhang /*109*/MatMatSolve_Elemental,
103840d92e34SHong Zhang        0,
103940d92e34SHong Zhang        0,
104040d92e34SHong Zhang        0,
104140d92e34SHong Zhang        0,
104240d92e34SHong Zhang /*114*/0,
104340d92e34SHong Zhang        0,
104440d92e34SHong Zhang        0,
104540d92e34SHong Zhang        0,
104640d92e34SHong Zhang        0,
104740d92e34SHong Zhang /*119*/0,
10484a29722dSXuan Zhou        MatHermitianTranspose_Elemental,
104940d92e34SHong Zhang        0,
105040d92e34SHong Zhang        0,
105140d92e34SHong Zhang        0,
105240d92e34SHong Zhang /*124*/0,
105340d92e34SHong Zhang        0,
105440d92e34SHong Zhang        0,
105540d92e34SHong Zhang        0,
105640d92e34SHong Zhang        0,
105740d92e34SHong Zhang /*129*/0,
105840d92e34SHong Zhang        0,
105940d92e34SHong Zhang        0,
106040d92e34SHong Zhang        0,
106140d92e34SHong Zhang        0,
106240d92e34SHong Zhang /*134*/0,
106340d92e34SHong Zhang        0,
106440d92e34SHong Zhang        0,
106540d92e34SHong Zhang        0,
106640d92e34SHong Zhang        0
106740d92e34SHong Zhang };
106840d92e34SHong Zhang 
1069ed36708cSHong Zhang /*MC
1070ed36708cSHong Zhang    MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1071ed36708cSHong Zhang 
1072ed36708cSHong Zhang    Options Database Keys:
10735cc86fc1SJed Brown + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
10745cc86fc1SJed Brown - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1075ed36708cSHong Zhang 
1076ed36708cSHong Zhang   Level: beginner
1077ed36708cSHong Zhang 
10785cb544a0SHong Zhang .seealso: MATDENSE
1079ed36708cSHong Zhang M*/
10804a29722dSXuan Zhou 
1081db31f6deSJed Brown #undef __FUNCT__
1082db31f6deSJed Brown #define __FUNCT__ "MatCreate_Elemental"
10838cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1084db31f6deSJed Brown {
1085db31f6deSJed Brown   Mat_Elemental      *a;
1086db31f6deSJed Brown   PetscErrorCode     ierr;
10875682a260SJack Poulson   PetscBool          flg,flg1;
10885e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
10895e9f5b67SHong Zhang   MPI_Comm           icomm;
10905682a260SJack Poulson   PetscInt           optv1;
1091db31f6deSJed Brown 
1092db31f6deSJed Brown   PetscFunctionBegin;
1093607a6623SBarry Smith   ierr = PetscElementalInitializePackage();CHKERRQ(ierr);
109440d92e34SHong Zhang   ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
109540d92e34SHong Zhang   A->insertmode = NOT_SET_VALUES;
1096db31f6deSJed Brown 
1097*b00a9115SJed Brown   ierr = PetscNewLog(A,&a);CHKERRQ(ierr);
1098db31f6deSJed Brown   A->data = (void*)a;
1099db31f6deSJed Brown 
1100db31f6deSJed Brown   /* Set up the elemental matrix */
1101ce94432eSBarry Smith   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
11025e9f5b67SHong Zhang 
11035e9f5b67SHong Zhang   /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
11045e9f5b67SHong Zhang   if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
1105180a43e4SHong Zhang     ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);
11065e9f5b67SHong Zhang   }
11070298fd71SBarry Smith   ierr = PetscCommDuplicate(cxxcomm,&icomm,NULL);CHKERRQ(ierr);
11085e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
11095e9f5b67SHong Zhang   if (!flg) {
1110*b00a9115SJed Brown     ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr);
11115cb544a0SHong Zhang 
1112ce94432eSBarry Smith     ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr);
11135cb544a0SHong Zhang     /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until elem::Grid() is called */
11142ef0cf24SXuan Zhou     ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",elem::mpi::CommSize(cxxcomm),&optv1,&flg1);CHKERRQ(ierr);
11155682a260SJack Poulson     if (flg1) {
11165682a260SJack Poulson       if (elem::mpi::CommSize(cxxcomm) % optv1 != 0) {
11175682a260SJack Poulson         SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)elem::mpi::CommSize(cxxcomm));
1118ed667823SXuan Zhou       }
11195682a260SJack Poulson       commgrid->grid = new elem::Grid(cxxcomm,optv1); /* use user-provided grid height */
11202ef0cf24SXuan Zhou     } else {
11212adf0be3SHong Zhang       commgrid->grid = new elem::Grid(cxxcomm); /* use Elemental default grid sizes */
1122ed667823SXuan Zhou     }
11235e9f5b67SHong Zhang     commgrid->grid_refct = 1;
11245e9f5b67SHong Zhang     ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr);
11255cb544a0SHong Zhang     PetscOptionsEnd();
11265e9f5b67SHong Zhang   } else {
11275e9f5b67SHong Zhang     commgrid->grid_refct++;
11285e9f5b67SHong Zhang   }
11295e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
11305e9f5b67SHong Zhang   a->grid      = commgrid->grid;
1131df311e6cSXuan Zhou   a->emat      = new elem::DistMatrix<PetscElemScalar>(*a->grid);
1132df311e6cSXuan Zhou   a->esubmat   = new elem::Matrix<PetscElemScalar>(1,1);
1133df311e6cSXuan Zhou   a->interface = new elem::AxpyInterface<PetscElemScalar>;
11347c920d81SXuan Zhou   a->pivot     = new elem::DistMatrix<PetscInt,elem::VC,elem::STAR>;
1135db31f6deSJed Brown 
1136db31f6deSJed Brown   /* build cache for off array entries formed */
1137aae2c449SHong Zhang   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
1138bafd5131SHong Zhang 
1139bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr);
1140bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetFactor_elemental_C",MatGetFactor_elemental_elemental);CHKERRQ(ierr);
1141db31f6deSJed Brown 
1142db31f6deSJed Brown   ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr);
1143db31f6deSJed Brown   PetscFunctionReturn(0);
1144db31f6deSJed Brown }
11454a29722dSXuan Zhou 
1146