xref: /petsc/src/mat/impls/elemental/matelem.cxx (revision e05a9d59ae70abb3b0c764000e407f332a4efdca)
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;
25e8146892SSatish Balay   if (El::Initialized()) PetscFunctionReturn(0);
26*e05a9d59SSatish Balay   El::Initialize();   /* called by the 1st call of MatCreate_Elemental */
27db31f6deSJed Brown   ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr);
28db31f6deSJed Brown   PetscFunctionReturn(0);
29db31f6deSJed Brown }
30db31f6deSJed Brown 
31db31f6deSJed Brown #undef __FUNCT__
32db31f6deSJed Brown #define __FUNCT__ "PetscElementalFinalizePackage"
33db31f6deSJed Brown /*@C
34db31f6deSJed Brown    PetscElementalFinalizePackage - Finalize Elemental package
35db31f6deSJed Brown 
36db31f6deSJed Brown    Logically Collective
37db31f6deSJed Brown 
38db31f6deSJed Brown    Level: developer
39db31f6deSJed Brown 
40db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage()
41db31f6deSJed Brown @*/
42db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void)
43db31f6deSJed Brown {
44db31f6deSJed Brown   PetscFunctionBegin;
45e8146892SSatish Balay   El::Finalize();  /* called by PetscFinalize() */
46db31f6deSJed Brown   PetscFunctionReturn(0);
47db31f6deSJed Brown }
48db31f6deSJed Brown 
49db31f6deSJed Brown #undef __FUNCT__
50db31f6deSJed Brown #define __FUNCT__ "MatView_Elemental"
51db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
52db31f6deSJed Brown {
53db31f6deSJed Brown   PetscErrorCode ierr;
54db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
55db31f6deSJed Brown   PetscBool      iascii;
56db31f6deSJed Brown 
57db31f6deSJed Brown   PetscFunctionBegin;
58db31f6deSJed Brown   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
59db31f6deSJed Brown   if (iascii) {
60db31f6deSJed Brown     PetscViewerFormat format;
61db31f6deSJed Brown     ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
62db31f6deSJed Brown     if (format == PETSC_VIEWER_ASCII_INFO) {
6379673f7bSHong Zhang       /* call elemental viewing function */
642d8adcc7SHong Zhang       ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr);
65ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr);
66ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr);
674fe7bbcaSHong Zhang       if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
6879673f7bSHong Zhang         /* call elemental viewing function */
69ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr);
704fe7bbcaSHong Zhang       }
7179673f7bSHong Zhang 
72db31f6deSJed Brown     } else if (format == PETSC_VIEWER_DEFAULT) {
73db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr);
74e8146892SSatish Balay       El::Print( *a->emat, "Elemental matrix (cyclic ordering)" );
75db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr);
76834d3fecSHong Zhang       if (A->factortype == MAT_FACTOR_NONE){
7761119200SXuan Zhou         Mat Adense;
78ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
7961119200SXuan Zhou         ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
8061119200SXuan Zhou         ierr = MatView(Adense,viewer);CHKERRQ(ierr);
8161119200SXuan Zhou         ierr = MatDestroy(&Adense);CHKERRQ(ierr);
82834d3fecSHong Zhang       }
83ce94432eSBarry Smith     } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
84d2daa67eSHong Zhang   } else {
855cb544a0SHong Zhang     /* convert to dense format and call MatView() */
8661119200SXuan Zhou     Mat Adense;
87ce94432eSBarry Smith     ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
8861119200SXuan Zhou     ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
8961119200SXuan Zhou     ierr = MatView(Adense,viewer);CHKERRQ(ierr);
9061119200SXuan Zhou     ierr = MatDestroy(&Adense);CHKERRQ(ierr);
91d2daa67eSHong Zhang   }
92db31f6deSJed Brown   PetscFunctionReturn(0);
93db31f6deSJed Brown }
94db31f6deSJed Brown 
95db31f6deSJed Brown #undef __FUNCT__
96180a43e4SHong Zhang #define __FUNCT__ "MatGetInfo_Elemental"
9715767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
98180a43e4SHong Zhang {
9915767789SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
10015767789SHong Zhang   PetscMPIInt    rank;
10115767789SHong Zhang 
102180a43e4SHong Zhang   PetscFunctionBegin;
103ce94432eSBarry Smith   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
10415767789SHong Zhang 
10515767789SHong Zhang   /* if (!rank) printf("          .........MatGetInfo_Elemental ...\n"); */
1065cb544a0SHong Zhang   info->block_size     = 1.0;
10715767789SHong Zhang 
10815767789SHong Zhang   if (flag == MAT_LOCAL) {
10915767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
11015767789SHong Zhang     info->nz_used        = info->nz_allocated;
11115767789SHong Zhang   } else if (flag == MAT_GLOBAL_MAX) {
112b2566f29SBarry Smith     //ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr);
11315767789SHong Zhang     /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
11415767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
11515767789SHong Zhang   } else if (flag == MAT_GLOBAL_SUM) {
11615767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
11715767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
11815767789SHong Zhang     info->nz_used        = info->nz_allocated; /* assume Elemental does accurate allocation */
119b2566f29SBarry Smith     //ierr = MPIU_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr);
12015767789SHong Zhang     //PetscPrintf(PETSC_COMM_SELF,"    ... [%d] locally allocated %g\n",rank,info->nz_allocated);
12115767789SHong Zhang   }
12215767789SHong Zhang 
12315767789SHong Zhang   info->nz_unneeded       = 0.0;
12415767789SHong Zhang   info->assemblies        = (double)A->num_ass;
12515767789SHong Zhang   info->mallocs           = 0;
12615767789SHong Zhang   info->memory            = ((PetscObject)A)->mem;
12715767789SHong Zhang   info->fill_ratio_given  = 0; /* determined by Elemental */
12815767789SHong Zhang   info->fill_ratio_needed = 0;
12915767789SHong Zhang   info->factor_mallocs    = 0;
130180a43e4SHong Zhang   PetscFunctionReturn(0);
131180a43e4SHong Zhang }
132180a43e4SHong Zhang 
133180a43e4SHong Zhang #undef __FUNCT__
134db31f6deSJed Brown #define __FUNCT__ "MatSetValues_Elemental"
135e6dea9dbSXuan Zhou static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode)
136db31f6deSJed Brown {
137db31f6deSJed Brown   PetscErrorCode ierr;
138db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
139db31f6deSJed Brown   PetscMPIInt    rank;
140db31f6deSJed Brown   PetscInt       i,j,rrank,ridx,crank,cidx;
141db31f6deSJed Brown 
142db31f6deSJed Brown   PetscFunctionBegin;
143ce94432eSBarry Smith   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr);
144db31f6deSJed Brown 
145e8146892SSatish Balay   const El::Grid &grid = a->emat->Grid();
146db31f6deSJed Brown   for (i=0; i<nr; i++) {
147db31f6deSJed Brown     PetscInt erow,ecol,elrow,elcol;
148db31f6deSJed Brown     if (rows[i] < 0) continue;
149db31f6deSJed Brown     P2RO(A,0,rows[i],&rrank,&ridx);
150db31f6deSJed Brown     RO2E(A,0,rrank,ridx,&erow);
151ce94432eSBarry Smith     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
152db31f6deSJed Brown     for (j=0; j<nc; j++) {
153db31f6deSJed Brown       if (cols[j] < 0) continue;
154db31f6deSJed Brown       P2RO(A,1,cols[j],&crank,&cidx);
155db31f6deSJed Brown       RO2E(A,1,crank,cidx,&ecol);
156ce94432eSBarry Smith       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
157aae2c449SHong Zhang       if (erow % grid.MCSize() != grid.MCRank() || ecol % grid.MRSize() != grid.MRRank()){ /* off-proc entry */
158aae2c449SHong Zhang         if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
159aae2c449SHong 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); */
160e6dea9dbSXuan Zhou         a->esubmat->Set(0,0, (PetscElemScalar)vals[i*nc+j]);
161aae2c449SHong Zhang         a->interface->Axpy(1.0,*(a->esubmat),erow,ecol);
162aae2c449SHong Zhang         continue;
163ed36708cSHong Zhang       }
164db31f6deSJed Brown       elrow = erow / grid.MCSize();
165db31f6deSJed Brown       elcol = ecol / grid.MRSize();
166db31f6deSJed Brown       switch (imode) {
167e6dea9dbSXuan Zhou       case INSERT_VALUES: a->emat->SetLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
168e6dea9dbSXuan Zhou       case ADD_VALUES: a->emat->UpdateLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
169ce94432eSBarry Smith       default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
170db31f6deSJed Brown       }
171db31f6deSJed Brown     }
172db31f6deSJed Brown   }
173db31f6deSJed Brown   PetscFunctionReturn(0);
174db31f6deSJed Brown }
175db31f6deSJed Brown 
176db31f6deSJed Brown #undef __FUNCT__
177db31f6deSJed Brown #define __FUNCT__ "MatMult_Elemental"
178db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
179db31f6deSJed Brown {
180db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
181db31f6deSJed Brown   PetscErrorCode        ierr;
182e6dea9dbSXuan Zhou   const PetscElemScalar *x;
183e6dea9dbSXuan Zhou   PetscElemScalar       *y;
184df311e6cSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
185db31f6deSJed Brown 
186db31f6deSJed Brown   PetscFunctionBegin;
187e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
188e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
189db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
190e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
1910c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
1920c18141cSBarry Smith     ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n);
193e8146892SSatish Balay     El::Gemv(El::NORMAL,one,*a->emat,xe,zero,ye);
194db31f6deSJed Brown   }
195e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
196e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
197db31f6deSJed Brown   PetscFunctionReturn(0);
198db31f6deSJed Brown }
199db31f6deSJed Brown 
200db31f6deSJed Brown #undef __FUNCT__
2019426833fSXuan Zhou #define __FUNCT__ "MatMultTranspose_Elemental"
2029426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
2039426833fSXuan Zhou {
2049426833fSXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
2059426833fSXuan Zhou   PetscErrorCode        ierr;
206df311e6cSXuan Zhou   const PetscElemScalar *x;
207df311e6cSXuan Zhou   PetscElemScalar       *y;
208e6dea9dbSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
2099426833fSXuan Zhou 
2109426833fSXuan Zhou   PetscFunctionBegin;
211e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
212e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2139426833fSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
214e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye;
2150c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
2160c18141cSBarry Smith     ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n);
217e8146892SSatish Balay     El::Gemv(El::TRANSPOSE,one,*a->emat,xe,zero,ye);
2189426833fSXuan Zhou   }
219e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
220e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2219426833fSXuan Zhou   PetscFunctionReturn(0);
2229426833fSXuan Zhou }
2239426833fSXuan Zhou 
2249426833fSXuan Zhou #undef __FUNCT__
225db31f6deSJed Brown #define __FUNCT__ "MatMultAdd_Elemental"
226db31f6deSJed Brown static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
227db31f6deSJed Brown {
228db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
229db31f6deSJed Brown   PetscErrorCode        ierr;
230df311e6cSXuan Zhou   const PetscElemScalar *x;
231df311e6cSXuan Zhou   PetscElemScalar       *z;
232e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
233db31f6deSJed Brown 
234db31f6deSJed Brown   PetscFunctionBegin;
235db31f6deSJed Brown   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
236e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
237e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
238db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
239e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
2400c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
2410c18141cSBarry Smith     ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n);
242e8146892SSatish Balay     El::Gemv(El::NORMAL,one,*a->emat,xe,one,ze);
243db31f6deSJed Brown   }
244e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
245e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
246db31f6deSJed Brown   PetscFunctionReturn(0);
247db31f6deSJed Brown }
248db31f6deSJed Brown 
249db31f6deSJed Brown #undef __FUNCT__
250e883f9d5SXuan Zhou #define __FUNCT__ "MatMultTransposeAdd_Elemental"
251e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
252e883f9d5SXuan Zhou {
253e883f9d5SXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
254e883f9d5SXuan Zhou   PetscErrorCode        ierr;
255df311e6cSXuan Zhou   const PetscElemScalar *x;
256df311e6cSXuan Zhou   PetscElemScalar       *z;
257e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
258e883f9d5SXuan Zhou 
259e883f9d5SXuan Zhou   PetscFunctionBegin;
260e883f9d5SXuan Zhou   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
261e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
262e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
263e883f9d5SXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
264e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze;
2650c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
2660c18141cSBarry Smith     ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n);
267e8146892SSatish Balay     El::Gemv(El::TRANSPOSE,one,*a->emat,xe,one,ze);
268e883f9d5SXuan Zhou   }
269e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
270e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
271e883f9d5SXuan Zhou   PetscFunctionReturn(0);
272e883f9d5SXuan Zhou }
273e883f9d5SXuan Zhou 
274e883f9d5SXuan Zhou #undef __FUNCT__
2759a9e8502SHong Zhang #define __FUNCT__ "MatMatMultNumeric_Elemental"
2769a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
277c1d1b975SXuan Zhou {
278c1d1b975SXuan Zhou   Mat_Elemental    *a = (Mat_Elemental*)A->data;
279c1d1b975SXuan Zhou   Mat_Elemental    *b = (Mat_Elemental*)B->data;
2809a9e8502SHong Zhang   Mat_Elemental    *c = (Mat_Elemental*)C->data;
281e6dea9dbSXuan Zhou   PetscElemScalar  one = 1,zero = 0;
282c1d1b975SXuan Zhou 
283c1d1b975SXuan Zhou   PetscFunctionBegin;
284aae2c449SHong Zhang   { /* Scoping so that constructor is called before pointer is returned */
285e8146892SSatish Balay     El::Gemm(El::NORMAL,El::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
286aae2c449SHong Zhang   }
2879a9e8502SHong Zhang   C->assembled = PETSC_TRUE;
2889a9e8502SHong Zhang   PetscFunctionReturn(0);
2899a9e8502SHong Zhang }
2909a9e8502SHong Zhang 
2919a9e8502SHong Zhang #undef __FUNCT__
2929a9e8502SHong Zhang #define __FUNCT__ "MatMatMultSymbolic_Elemental"
2939a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
2949a9e8502SHong Zhang {
2959a9e8502SHong Zhang   PetscErrorCode ierr;
2969a9e8502SHong Zhang   Mat            Ce;
297ce94432eSBarry Smith   MPI_Comm       comm;
2989a9e8502SHong Zhang 
2999a9e8502SHong Zhang   PetscFunctionBegin;
300ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
3019a9e8502SHong Zhang   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
3029a9e8502SHong Zhang   ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
3039a9e8502SHong Zhang   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
3049a9e8502SHong Zhang   ierr = MatSetUp(Ce);CHKERRQ(ierr);
3059a9e8502SHong Zhang   *C = Ce;
3069a9e8502SHong Zhang   PetscFunctionReturn(0);
3079a9e8502SHong Zhang }
3089a9e8502SHong Zhang 
3099a9e8502SHong Zhang #undef __FUNCT__
3109a9e8502SHong Zhang #define __FUNCT__ "MatMatMult_Elemental"
3119a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3129a9e8502SHong Zhang {
3139a9e8502SHong Zhang   PetscErrorCode ierr;
3149a9e8502SHong Zhang 
3159a9e8502SHong Zhang   PetscFunctionBegin;
3169a9e8502SHong Zhang   if (scall == MAT_INITIAL_MATRIX){
3173ff4c91cSHong Zhang     ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3189a9e8502SHong Zhang     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
3193ff4c91cSHong Zhang     ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3209a9e8502SHong Zhang   }
3213ff4c91cSHong Zhang   ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
3229a9e8502SHong Zhang   ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
3233ff4c91cSHong Zhang   ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
324c1d1b975SXuan Zhou   PetscFunctionReturn(0);
325c1d1b975SXuan Zhou }
326c1d1b975SXuan Zhou 
327c1d1b975SXuan Zhou #undef __FUNCT__
328df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultNumeric_Elemental"
329df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
330df311e6cSXuan Zhou {
331df311e6cSXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
332df311e6cSXuan Zhou   Mat_Elemental      *b = (Mat_Elemental*)B->data;
333df311e6cSXuan Zhou   Mat_Elemental      *c = (Mat_Elemental*)C->data;
334e6dea9dbSXuan Zhou   PetscElemScalar    one = 1,zero = 0;
335df311e6cSXuan Zhou 
336df311e6cSXuan Zhou   PetscFunctionBegin;
337df311e6cSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
338e8146892SSatish Balay     El::Gemm(El::NORMAL,El::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
339df311e6cSXuan Zhou   }
340df311e6cSXuan Zhou   C->assembled = PETSC_TRUE;
341df311e6cSXuan Zhou   PetscFunctionReturn(0);
342df311e6cSXuan Zhou }
343df311e6cSXuan Zhou 
344df311e6cSXuan Zhou #undef __FUNCT__
345df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultSymbolic_Elemental"
346df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
347df311e6cSXuan Zhou {
348df311e6cSXuan Zhou   PetscErrorCode ierr;
349df311e6cSXuan Zhou   Mat            Ce;
350ce94432eSBarry Smith   MPI_Comm       comm;
351df311e6cSXuan Zhou 
352df311e6cSXuan Zhou   PetscFunctionBegin;
353ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
354df311e6cSXuan Zhou   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
355df311e6cSXuan Zhou   ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
356df311e6cSXuan Zhou   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
357df311e6cSXuan Zhou   ierr = MatSetUp(Ce);CHKERRQ(ierr);
358df311e6cSXuan Zhou   *C = Ce;
359df311e6cSXuan Zhou   PetscFunctionReturn(0);
360df311e6cSXuan Zhou }
361df311e6cSXuan Zhou 
362df311e6cSXuan Zhou #undef __FUNCT__
363df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMult_Elemental"
364df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
365df311e6cSXuan Zhou {
366df311e6cSXuan Zhou   PetscErrorCode ierr;
367df311e6cSXuan Zhou 
368df311e6cSXuan Zhou   PetscFunctionBegin;
369df311e6cSXuan Zhou   if (scall == MAT_INITIAL_MATRIX){
370df311e6cSXuan Zhou     ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
371df311e6cSXuan Zhou     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
372df311e6cSXuan Zhou     ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
373df311e6cSXuan Zhou   }
374df311e6cSXuan Zhou   ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
375df311e6cSXuan Zhou   ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
376df311e6cSXuan Zhou   ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
377df311e6cSXuan Zhou   PetscFunctionReturn(0);
378df311e6cSXuan Zhou }
379df311e6cSXuan Zhou 
380df311e6cSXuan Zhou #undef __FUNCT__
38161119200SXuan Zhou #define __FUNCT__ "MatGetDiagonal_Elemental"
382a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
38361119200SXuan Zhou {
384a9d89745SXuan Zhou   PetscInt        i,nrows,ncols,nD,rrank,ridx,crank,cidx;
385a9d89745SXuan Zhou   Mat_Elemental   *a = (Mat_Elemental*)A->data;
38661119200SXuan Zhou   PetscErrorCode  ierr;
387a9d89745SXuan Zhou   PetscElemScalar v;
388ce94432eSBarry Smith   MPI_Comm        comm;
38961119200SXuan Zhou 
39061119200SXuan Zhou   PetscFunctionBegin;
391ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
392a9d89745SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
393a9d89745SXuan Zhou   nD = nrows>ncols ? ncols : nrows;
394a9d89745SXuan Zhou   for (i=0; i<nD; i++) {
395a9d89745SXuan Zhou     PetscInt erow,ecol;
396a9d89745SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
397a9d89745SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
398a9d89745SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
399a9d89745SXuan Zhou     P2RO(A,1,i,&crank,&cidx);
400a9d89745SXuan Zhou     RO2E(A,1,crank,cidx,&ecol);
401a9d89745SXuan Zhou     if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
402a9d89745SXuan Zhou     v = a->emat->Get(erow,ecol);
4037c8b904fSSatish Balay     ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr);
404ade3cc5eSXuan Zhou   }
405a9d89745SXuan Zhou   ierr = VecAssemblyBegin(D);CHKERRQ(ierr);
406a9d89745SXuan Zhou   ierr = VecAssemblyEnd(D);CHKERRQ(ierr);
40761119200SXuan Zhou   PetscFunctionReturn(0);
40861119200SXuan Zhou }
40961119200SXuan Zhou 
41061119200SXuan Zhou #undef __FUNCT__
411ade3cc5eSXuan Zhou #define __FUNCT__ "MatDiagonalScale_Elemental"
412ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
413ade3cc5eSXuan Zhou {
414ade3cc5eSXuan Zhou   Mat_Elemental         *x = (Mat_Elemental*)X->data;
415ade3cc5eSXuan Zhou   const PetscElemScalar *d;
416ade3cc5eSXuan Zhou   PetscErrorCode        ierr;
417ade3cc5eSXuan Zhou 
418ade3cc5eSXuan Zhou   PetscFunctionBegin;
4199065cd98SJed Brown   if (R) {
420ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
421e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
4220c18141cSBarry Smith     de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n);
423e8146892SSatish Balay     El::DiagonalScale(El::RIGHT,El::NORMAL,de,*x->emat);
424ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
4259065cd98SJed Brown   }
4269065cd98SJed Brown   if (L) {
427ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
428e8146892SSatish Balay     El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de;
4290c18141cSBarry Smith     de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n);
430e8146892SSatish Balay     El::DiagonalScale(El::LEFT,El::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;
443e8146892SSatish Balay   El::Scale((PetscElemScalar)a,*x->emat);
44465b78793SXuan Zhou   PetscFunctionReturn(0);
44565b78793SXuan Zhou }
44665b78793SXuan Zhou 
447f82baa17SHong Zhang /*
448f82baa17SHong Zhang   MatAXPY - Computes Y = a*X + Y.
449f82baa17SHong Zhang */
45065b78793SXuan Zhou #undef __FUNCT__
4514ee44ac6SXuan Zhou #define __FUNCT__ "MatAXPY_Elemental"
452e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
453e09a3074SHong Zhang {
454e09a3074SHong Zhang   Mat_Elemental  *x = (Mat_Elemental*)X->data;
455e09a3074SHong Zhang   Mat_Elemental  *y = (Mat_Elemental*)Y->data;
45668446cf8SJed Brown   PetscErrorCode ierr;
457e09a3074SHong Zhang 
458e09a3074SHong Zhang   PetscFunctionBegin;
459e8146892SSatish Balay   El::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
46021f6c9c4SJed Brown   ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr);
461e09a3074SHong Zhang   PetscFunctionReturn(0);
462e09a3074SHong Zhang }
463e09a3074SHong Zhang 
464ae844d54SHong Zhang #undef __FUNCT__
465d6223691SXuan Zhou #define __FUNCT__ "MatCopy_Elemental"
466d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
467d6223691SXuan Zhou {
468d6223691SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
469d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
470d6223691SXuan Zhou 
471d6223691SXuan Zhou   PetscFunctionBegin;
472e8146892SSatish Balay   El::Copy(*a->emat,*b->emat);
473d6223691SXuan Zhou   PetscFunctionReturn(0);
474d6223691SXuan Zhou }
475d6223691SXuan Zhou 
476d6223691SXuan Zhou #undef __FUNCT__
477df311e6cSXuan Zhou #define __FUNCT__ "MatDuplicate_Elemental"
478df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
479df311e6cSXuan Zhou {
480df311e6cSXuan Zhou   Mat            Be;
481ce94432eSBarry Smith   MPI_Comm       comm;
482df311e6cSXuan Zhou   Mat_Elemental  *a=(Mat_Elemental*)A->data;
483df311e6cSXuan Zhou   PetscErrorCode ierr;
484df311e6cSXuan Zhou 
485df311e6cSXuan Zhou   PetscFunctionBegin;
486ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
487df311e6cSXuan Zhou   ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
488df311e6cSXuan Zhou   ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
489df311e6cSXuan Zhou   ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
490df311e6cSXuan Zhou   ierr = MatSetUp(Be);CHKERRQ(ierr);
491df311e6cSXuan Zhou   *B = Be;
492df311e6cSXuan Zhou   if (op == MAT_COPY_VALUES) {
493df311e6cSXuan Zhou     Mat_Elemental *b=(Mat_Elemental*)Be->data;
494e8146892SSatish Balay     El::Copy(*a->emat,*b->emat);
495df311e6cSXuan Zhou   }
496df311e6cSXuan Zhou   Be->assembled = PETSC_TRUE;
497df311e6cSXuan Zhou   PetscFunctionReturn(0);
498df311e6cSXuan Zhou }
499df311e6cSXuan Zhou 
500df311e6cSXuan Zhou #undef __FUNCT__
501d6223691SXuan Zhou #define __FUNCT__ "MatTranspose_Elemental"
502d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
503d6223691SXuan Zhou {
504ec8cb81fSBarry Smith   Mat            Be = *B;
5055262d616SXuan Zhou   PetscErrorCode ierr;
506ce94432eSBarry Smith   MPI_Comm       comm;
5074fe7bbcaSHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
508d6223691SXuan Zhou 
509d6223691SXuan Zhou   PetscFunctionBegin;
510ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5115cb544a0SHong Zhang   /* Only out-of-place supported */
5125262d616SXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5135262d616SXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5145262d616SXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5155262d616SXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5165262d616SXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5175262d616SXuan Zhou     *B = Be;
5185262d616SXuan Zhou   }
5194fe7bbcaSHong Zhang   b = (Mat_Elemental*)Be->data;
520e8146892SSatish Balay   El::Transpose(*a->emat,*b->emat);
5215262d616SXuan Zhou   Be->assembled = PETSC_TRUE;
522d6223691SXuan Zhou   PetscFunctionReturn(0);
523d6223691SXuan Zhou }
524d6223691SXuan Zhou 
525d6223691SXuan Zhou #undef __FUNCT__
526dfcb0403SXuan Zhou #define __FUNCT__ "MatConjugate_Elemental"
527dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A)
528dfcb0403SXuan Zhou {
529dfcb0403SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
530dfcb0403SXuan Zhou 
531dfcb0403SXuan Zhou   PetscFunctionBegin;
532e8146892SSatish Balay   El::Conjugate(*a->emat);
533dfcb0403SXuan Zhou   PetscFunctionReturn(0);
534dfcb0403SXuan Zhou }
535dfcb0403SXuan Zhou 
536dfcb0403SXuan Zhou #undef __FUNCT__
5374a29722dSXuan Zhou #define __FUNCT__ "MatHermitianTranspose_Elemental"
5384a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
5394a29722dSXuan Zhou {
540ec8cb81fSBarry Smith   Mat            Be = *B;
5414a29722dSXuan Zhou   PetscErrorCode ierr;
542ce94432eSBarry Smith   MPI_Comm       comm;
5434a29722dSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
5444a29722dSXuan Zhou 
5454a29722dSXuan Zhou   PetscFunctionBegin;
546ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5475cb544a0SHong Zhang   /* Only out-of-place supported */
5484a29722dSXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5494a29722dSXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5504a29722dSXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5514a29722dSXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5524a29722dSXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5534a29722dSXuan Zhou     *B = Be;
5544a29722dSXuan Zhou   }
5554a29722dSXuan Zhou   b = (Mat_Elemental*)Be->data;
556e8146892SSatish Balay   El::Adjoint(*a->emat,*b->emat);
5574a29722dSXuan Zhou   Be->assembled = PETSC_TRUE;
5584a29722dSXuan Zhou   PetscFunctionReturn(0);
5594a29722dSXuan Zhou }
5604a29722dSXuan Zhou 
5614a29722dSXuan Zhou #undef __FUNCT__
5621f881ff8SXuan Zhou #define __FUNCT__ "MatSolve_Elemental"
5631f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
5641f881ff8SXuan Zhou {
5651f881ff8SXuan Zhou   Mat_Elemental     *a = (Mat_Elemental*)A->data;
5661f881ff8SXuan Zhou   PetscErrorCode    ierr;
567df311e6cSXuan Zhou   PetscElemScalar   *x;
5681f881ff8SXuan Zhou 
5691f881ff8SXuan Zhou   PetscFunctionBegin;
57045cf121fSXuan Zhou   ierr = VecCopy(B,X);CHKERRQ(ierr);
571e6dea9dbSXuan Zhou   ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
572e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe;
5730c18141cSBarry Smith   xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
574e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::MC,El::MR> xer(xe);
575fc54b460SXuan Zhou   switch (A->factortype) {
576fc54b460SXuan Zhou   case MAT_FACTOR_LU:
5771f881ff8SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
578e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->pivot,xer);
579e8146892SSatish Balay       El::Copy(xer,xe);
580fc54b460SXuan Zhou     } else {
581e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,xer);
582e8146892SSatish Balay       El::Copy(xer,xe);
5831f881ff8SXuan Zhou     }
584fc54b460SXuan Zhou     break;
585fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
586e8146892SSatish Balay     El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,xer);
587e8146892SSatish Balay     El::Copy(xer,xe);
588fc54b460SXuan Zhou     break;
589fc54b460SXuan Zhou   default:
5904fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
591fc54b460SXuan Zhou     break;
5921f881ff8SXuan Zhou   }
593e6dea9dbSXuan Zhou   ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
5941f881ff8SXuan Zhou   PetscFunctionReturn(0);
5951f881ff8SXuan Zhou }
5961f881ff8SXuan Zhou 
5971f881ff8SXuan Zhou #undef __FUNCT__
598df311e6cSXuan Zhou #define __FUNCT__ "MatSolveAdd_Elemental"
599df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
600df311e6cSXuan Zhou {
601df311e6cSXuan Zhou   PetscErrorCode    ierr;
602df311e6cSXuan Zhou 
603df311e6cSXuan Zhou   PetscFunctionBegin;
604df311e6cSXuan Zhou   ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr);
6053d7f40dbSXuan Zhou   ierr = VecAXPY(X,1,Y);CHKERRQ(ierr);
606df311e6cSXuan Zhou   PetscFunctionReturn(0);
607df311e6cSXuan Zhou }
608df311e6cSXuan Zhou 
609df311e6cSXuan Zhou #undef __FUNCT__
610ae844d54SHong Zhang #define __FUNCT__ "MatMatSolve_Elemental"
611ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
612ae844d54SHong Zhang {
6131f0e42cfSHong Zhang   Mat_Elemental *a=(Mat_Elemental*)A->data;
614d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
6151f0e42cfSHong Zhang   Mat_Elemental *x=(Mat_Elemental*)X->data;
6161f0e42cfSHong Zhang 
617ae844d54SHong Zhang   PetscFunctionBegin;
618e8146892SSatish Balay   El::Copy(*b->emat,*x->emat);
619fc54b460SXuan Zhou   switch (A->factortype) {
620fc54b460SXuan Zhou   case MAT_FACTOR_LU:
621d6223691SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
622e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*a->pivot,*x->emat);
623fc54b460SXuan Zhou     } else {
624e8146892SSatish Balay       El::lu::SolveAfter(El::NORMAL,*a->emat,*x->emat);
625d6223691SXuan Zhou     }
626fc54b460SXuan Zhou     break;
627fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
628e8146892SSatish Balay     El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,*x->emat);
629fc54b460SXuan Zhou     break;
630fc54b460SXuan Zhou   default:
6314fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
632fc54b460SXuan Zhou     break;
633fc54b460SXuan Zhou   }
634ae844d54SHong Zhang   PetscFunctionReturn(0);
635ae844d54SHong Zhang }
636ae844d54SHong Zhang 
637ae844d54SHong Zhang #undef __FUNCT__
638ae844d54SHong Zhang #define __FUNCT__ "MatLUFactor_Elemental"
639ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
640ae844d54SHong Zhang {
6417c920d81SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
6427c920d81SXuan Zhou 
643ae844d54SHong Zhang   PetscFunctionBegin;
644d6223691SXuan Zhou   if (info->dtcol){
645e8146892SSatish Balay     El::LU(*a->emat,*a->pivot);
6461293973dSHong Zhang   } else {
647e8146892SSatish Balay     El::LU(*a->emat);
648d6223691SXuan Zhou   }
6491293973dSHong Zhang   A->factortype = MAT_FACTOR_LU;
650834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
651ae844d54SHong Zhang   PetscFunctionReturn(0);
652ae844d54SHong Zhang }
653ae844d54SHong Zhang 
654d7c3f9d8SHong Zhang #undef __FUNCT__
655d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorNumeric_Elemental"
656d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
657d7c3f9d8SHong Zhang {
658d7c3f9d8SHong Zhang   PetscErrorCode ierr;
659d7c3f9d8SHong Zhang 
660d7c3f9d8SHong Zhang   PetscFunctionBegin;
661d7c3f9d8SHong Zhang   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
662d7c3f9d8SHong Zhang   ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr);
663d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
664d7c3f9d8SHong Zhang }
665d7c3f9d8SHong Zhang 
666d7c3f9d8SHong Zhang #undef __FUNCT__
667d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorSymbolic_Elemental"
668d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
669d7c3f9d8SHong Zhang {
670d7c3f9d8SHong Zhang   PetscFunctionBegin;
671d7c3f9d8SHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
672d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
673d7c3f9d8SHong Zhang }
674d7c3f9d8SHong Zhang 
67545cf121fSXuan Zhou #undef __FUNCT__
67645cf121fSXuan Zhou #define __FUNCT__ "MatCholeskyFactor_Elemental"
67745cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
67845cf121fSXuan Zhou {
67945cf121fSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
680e8146892SSatish Balay   El::DistMatrix<PetscElemScalar,El::MC,El::STAR> d;
68145cf121fSXuan Zhou 
68245cf121fSXuan Zhou   PetscFunctionBegin;
683e8146892SSatish Balay   El::Cholesky(El::UPPER,*a->emat);
684fc54b460SXuan Zhou   A->factortype = MAT_FACTOR_CHOLESKY;
685834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
68645cf121fSXuan Zhou   PetscFunctionReturn(0);
68745cf121fSXuan Zhou }
68845cf121fSXuan Zhou 
68979673f7bSHong Zhang #undef __FUNCT__
69079673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorNumeric_Elemental"
69179673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
69279673f7bSHong Zhang {
693cb76c1d8SXuan Zhou   PetscErrorCode ierr;
694cb76c1d8SXuan Zhou 
695cb76c1d8SXuan Zhou   PetscFunctionBegin;
696cb76c1d8SXuan Zhou   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
697cb76c1d8SXuan Zhou   ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr);
698cb76c1d8SXuan Zhou   PetscFunctionReturn(0);
69979673f7bSHong Zhang }
700cb76c1d8SXuan Zhou 
70179673f7bSHong Zhang #undef __FUNCT__
70279673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorSymbolic_Elemental"
70379673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
70479673f7bSHong Zhang {
70579673f7bSHong Zhang   PetscFunctionBegin;
70679673f7bSHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
70779673f7bSHong Zhang   PetscFunctionReturn(0);
70879673f7bSHong Zhang }
70979673f7bSHong Zhang 
7101293973dSHong Zhang #undef __FUNCT__
71115767789SHong Zhang #define __FUNCT__ "MatFactorGetSolverPackage_elemental_elemental"
71215767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type)
71315767789SHong Zhang {
71415767789SHong Zhang   PetscFunctionBegin;
71515767789SHong Zhang   *type = MATSOLVERELEMENTAL;
71615767789SHong Zhang   PetscFunctionReturn(0);
71715767789SHong Zhang }
71815767789SHong Zhang 
71915767789SHong Zhang #undef __FUNCT__
72015767789SHong Zhang #define __FUNCT__ "MatGetFactor_elemental_elemental"
72115767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
7221293973dSHong Zhang {
7231293973dSHong Zhang   Mat            B;
7241293973dSHong Zhang   PetscErrorCode ierr;
7251293973dSHong Zhang 
7261293973dSHong Zhang   PetscFunctionBegin;
7271293973dSHong Zhang   /* Create the factorization matrix */
728ce94432eSBarry Smith   ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
7291293973dSHong Zhang   ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
7301293973dSHong Zhang   ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr);
7311293973dSHong Zhang   ierr = MatSetUp(B);CHKERRQ(ierr);
7321293973dSHong Zhang   B->factortype = ftype;
733bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr);
7341293973dSHong Zhang   *F            = B;
7351293973dSHong Zhang   PetscFunctionReturn(0);
7361293973dSHong Zhang }
7371293973dSHong Zhang 
7381f881ff8SXuan Zhou #undef __FUNCT__
73942c9c57cSBarry Smith #define __FUNCT__ "MatSolverPackageRegister_Elemental"
74029b38603SBarry Smith PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_Elemental(void)
74142c9c57cSBarry Smith {
74218713533SBarry Smith   PetscErrorCode ierr;
74318713533SBarry Smith 
74442c9c57cSBarry Smith   PetscFunctionBegin;
74542c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_LU,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
74642c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
74742c9c57cSBarry Smith   PetscFunctionReturn(0);
74842c9c57cSBarry Smith }
74942c9c57cSBarry Smith 
75042c9c57cSBarry Smith #undef __FUNCT__
7511f881ff8SXuan Zhou #define __FUNCT__ "MatNorm_Elemental"
7521f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
7531f881ff8SXuan Zhou {
7541f881ff8SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7551f881ff8SXuan Zhou 
7561f881ff8SXuan Zhou   PetscFunctionBegin;
7571f881ff8SXuan Zhou   switch (type){
7581f881ff8SXuan Zhou   case NORM_1:
759e8146892SSatish Balay     *nrm = El::OneNorm(*a->emat);
7601f881ff8SXuan Zhou     break;
7611f881ff8SXuan Zhou   case NORM_FROBENIUS:
762e8146892SSatish Balay     *nrm = El::FrobeniusNorm(*a->emat);
7631f881ff8SXuan Zhou     break;
7641f881ff8SXuan Zhou   case NORM_INFINITY:
765e8146892SSatish Balay     *nrm = El::InfinityNorm(*a->emat);
7661f881ff8SXuan Zhou     break;
7671f881ff8SXuan Zhou   default:
7681f881ff8SXuan Zhou     printf("Error: unsupported norm type!\n");
7691f881ff8SXuan Zhou   }
7701f881ff8SXuan Zhou   PetscFunctionReturn(0);
7711f881ff8SXuan Zhou }
7721f881ff8SXuan Zhou 
7735262d616SXuan Zhou #undef __FUNCT__
7745262d616SXuan Zhou #define __FUNCT__ "MatZeroEntries_Elemental"
7755262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A)
7765262d616SXuan Zhou {
7775262d616SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7785262d616SXuan Zhou 
7795262d616SXuan Zhou   PetscFunctionBegin;
780e8146892SSatish Balay   El::Zero(*a->emat);
7815262d616SXuan Zhou   PetscFunctionReturn(0);
7825262d616SXuan Zhou }
7835262d616SXuan Zhou 
784e09a3074SHong Zhang #undef __FUNCT__
785db31f6deSJed Brown #define __FUNCT__ "MatGetOwnershipIS_Elemental"
786db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
787db31f6deSJed Brown {
788db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
789db31f6deSJed Brown   PetscErrorCode ierr;
790db31f6deSJed Brown   PetscInt       i,m,shift,stride,*idx;
791db31f6deSJed Brown 
792db31f6deSJed Brown   PetscFunctionBegin;
793db31f6deSJed Brown   if (rows) {
794db31f6deSJed Brown     m = a->emat->LocalHeight();
795db31f6deSJed Brown     shift = a->emat->ColShift();
796db31f6deSJed Brown     stride = a->emat->ColStride();
797785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
798db31f6deSJed Brown     for (i=0; i<m; i++) {
799db31f6deSJed Brown       PetscInt rank,offset;
800db31f6deSJed Brown       E2RO(A,0,shift+i*stride,&rank,&offset);
801db31f6deSJed Brown       RO2P(A,0,rank,offset,&idx[i]);
802db31f6deSJed Brown     }
803db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
804db31f6deSJed Brown   }
805db31f6deSJed Brown   if (cols) {
806db31f6deSJed Brown     m = a->emat->LocalWidth();
807db31f6deSJed Brown     shift = a->emat->RowShift();
808db31f6deSJed Brown     stride = a->emat->RowStride();
809785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
810db31f6deSJed Brown     for (i=0; i<m; i++) {
811db31f6deSJed Brown       PetscInt rank,offset;
812db31f6deSJed Brown       E2RO(A,1,shift+i*stride,&rank,&offset);
813db31f6deSJed Brown       RO2P(A,1,rank,offset,&idx[i]);
814db31f6deSJed Brown     }
815db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr);
816db31f6deSJed Brown   }
817db31f6deSJed Brown   PetscFunctionReturn(0);
818db31f6deSJed Brown }
819db31f6deSJed Brown 
820db31f6deSJed Brown #undef __FUNCT__
8212ef0cf24SXuan Zhou #define __FUNCT__ "MatConvert_Elemental_Dense"
82219fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
823af295397SXuan Zhou {
8242ef0cf24SXuan Zhou   Mat                Bmpi;
825af295397SXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
826ce94432eSBarry Smith   MPI_Comm           comm;
8272ef0cf24SXuan Zhou   PetscErrorCode     ierr;
8282ef0cf24SXuan Zhou   PetscInt           rrank,ridx,crank,cidx,nrows,ncols,i,j;
829df311e6cSXuan Zhou   PetscElemScalar    v;
830573b0fb4SBarry Smith   PetscBool          s1,s2,s3;
831af295397SXuan Zhou 
832af295397SXuan Zhou   PetscFunctionBegin;
833ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
834573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATDENSE,&s1);CHKERRQ(ierr);
835573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATSEQDENSE,&s2);CHKERRQ(ierr);
836573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATMPIDENSE,&s3);CHKERRQ(ierr);
837573b0fb4SBarry Smith   if (!s1 && !s2 && !s3) SETERRQ(comm,PETSC_ERR_SUP,"Unsupported New MatType: must be MATDENSE, MATSEQDENSE or MATMPIDENSE");
838af295397SXuan Zhou   ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr);
839af295397SXuan Zhou   ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
8402ef0cf24SXuan Zhou   ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr);
841af295397SXuan Zhou   ierr = MatSetUp(Bmpi);CHKERRQ(ierr);
8422ef0cf24SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
8432ef0cf24SXuan Zhou   for (i=0; i<nrows; i++) {
8442ef0cf24SXuan Zhou     PetscInt erow,ecol;
8452ef0cf24SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
8462ef0cf24SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
8472ef0cf24SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
8482ef0cf24SXuan Zhou     for (j=0; j<ncols; j++) {
8492ef0cf24SXuan Zhou       P2RO(A,1,j,&crank,&cidx);
8502ef0cf24SXuan Zhou       RO2E(A,1,crank,cidx,&ecol);
8512ef0cf24SXuan Zhou       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
8522ef0cf24SXuan Zhou       v = a->emat->Get(erow,ecol);
853e6dea9dbSXuan Zhou       ierr = MatSetValues(Bmpi,1,&i,1,&j,(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
8542ef0cf24SXuan Zhou     }
8552ef0cf24SXuan Zhou   }
856af295397SXuan Zhou   ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
857af295397SXuan Zhou   ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
858c4ad791aSXuan Zhou   if (reuse == MAT_REUSE_MATRIX) {
85928be2f97SBarry Smith     ierr = MatHeaderReplace(A,&Bmpi);CHKERRQ(ierr);
860c4ad791aSXuan Zhou   } else {
861c4ad791aSXuan Zhou     *B = Bmpi;
862c4ad791aSXuan Zhou   }
863af295397SXuan Zhou   PetscFunctionReturn(0);
864af295397SXuan Zhou }
865af295397SXuan Zhou 
866af295397SXuan Zhou #undef __FUNCT__
867af8000cdSHong Zhang #define __FUNCT__ "MatConvert_SeqAIJ_Elemental"
868af8000cdSHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
869af8000cdSHong Zhang {
870af8000cdSHong Zhang   Mat               mat_elemental;
871af8000cdSHong Zhang   PetscErrorCode    ierr;
872af8000cdSHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols;
873af8000cdSHong Zhang   const PetscInt    *cols;
874af8000cdSHong Zhang   const PetscScalar *vals;
875af8000cdSHong Zhang 
876af8000cdSHong Zhang   PetscFunctionBegin;
877af8000cdSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
878af8000cdSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
879af8000cdSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
880af8000cdSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
881af8000cdSHong Zhang   for (row=0; row<M; row++) {
882af8000cdSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
883af8000cdSHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
884af8000cdSHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
885af8000cdSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
886af8000cdSHong Zhang   }
887af8000cdSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
888af8000cdSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
889af8000cdSHong Zhang 
890af8000cdSHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
89128be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
892af8000cdSHong Zhang   } else {
893af8000cdSHong Zhang     *newmat = mat_elemental;
894af8000cdSHong Zhang   }
895af8000cdSHong Zhang   PetscFunctionReturn(0);
896af8000cdSHong Zhang }
897af8000cdSHong Zhang 
898af8000cdSHong Zhang #undef __FUNCT__
8995d7652ecSHong Zhang #define __FUNCT__ "MatConvert_MPIAIJ_Elemental"
9005d7652ecSHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9015d7652ecSHong Zhang {
9025d7652ecSHong Zhang   Mat               mat_elemental;
9035d7652ecSHong Zhang   PetscErrorCode    ierr;
9045d7652ecSHong Zhang   PetscInt          row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j;
9055d7652ecSHong Zhang   const PetscInt    *cols;
9065d7652ecSHong Zhang   const PetscScalar *vals;
9075d7652ecSHong Zhang 
9085d7652ecSHong Zhang   PetscFunctionBegin;
9095d7652ecSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9105d7652ecSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);CHKERRQ(ierr);
9115d7652ecSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9125d7652ecSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9135d7652ecSHong Zhang   for (row=rstart; row<rend; row++) {
9145d7652ecSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9155d7652ecSHong Zhang     for (j=0; j<ncols; j++) {
9165d7652ecSHong Zhang       /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9175d7652ecSHong Zhang       ierr = MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);CHKERRQ(ierr);
9185d7652ecSHong Zhang     }
9195d7652ecSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9205d7652ecSHong Zhang   }
9215d7652ecSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9225d7652ecSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9235d7652ecSHong Zhang 
9245d7652ecSHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
92528be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
9265d7652ecSHong Zhang   } else {
9275d7652ecSHong Zhang     *newmat = mat_elemental;
9285d7652ecSHong Zhang   }
9295d7652ecSHong Zhang   PetscFunctionReturn(0);
9305d7652ecSHong Zhang }
9315d7652ecSHong Zhang 
9325d7652ecSHong Zhang #undef __FUNCT__
9336214f412SHong Zhang #define __FUNCT__ "MatConvert_SeqSBAIJ_Elemental"
9346214f412SHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_SeqSBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9356214f412SHong Zhang {
9366214f412SHong Zhang   Mat               mat_elemental;
9376214f412SHong Zhang   PetscErrorCode    ierr;
9386214f412SHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols,j;
9396214f412SHong Zhang   const PetscInt    *cols;
9406214f412SHong Zhang   const PetscScalar *vals;
9416214f412SHong Zhang 
9426214f412SHong Zhang   PetscFunctionBegin;
9436214f412SHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9446214f412SHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
9456214f412SHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9466214f412SHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9476214f412SHong Zhang   ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr);
9486214f412SHong Zhang   for (row=0; row<M; row++) {
9496214f412SHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9506214f412SHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9516214f412SHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
9526214f412SHong Zhang     for (j=0; j<ncols; j++) { /* lower triangular part */
9536214f412SHong Zhang       if (cols[j] == row) continue;
9546214f412SHong Zhang       ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr);
9556214f412SHong Zhang     }
9566214f412SHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9576214f412SHong Zhang   }
9586214f412SHong Zhang   ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr);
9596214f412SHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9606214f412SHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9616214f412SHong Zhang 
9626214f412SHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
96328be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
9646214f412SHong Zhang   } else {
9656214f412SHong Zhang     *newmat = mat_elemental;
9666214f412SHong Zhang   }
9676214f412SHong Zhang   PetscFunctionReturn(0);
9686214f412SHong Zhang }
9696214f412SHong Zhang 
9706214f412SHong Zhang #undef __FUNCT__
9716214f412SHong Zhang #define __FUNCT__ "MatConvert_MPISBAIJ_Elemental"
9726214f412SHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_MPISBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9736214f412SHong Zhang {
9746214f412SHong Zhang   Mat               mat_elemental;
9756214f412SHong Zhang   PetscErrorCode    ierr;
9766214f412SHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols,j,rstart=A->rmap->rstart,rend=A->rmap->rend;
9776214f412SHong Zhang   const PetscInt    *cols;
9786214f412SHong Zhang   const PetscScalar *vals;
9796214f412SHong Zhang 
9806214f412SHong Zhang   PetscFunctionBegin;
9816214f412SHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9826214f412SHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
9836214f412SHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9846214f412SHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9856214f412SHong Zhang   ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr);
9866214f412SHong Zhang   for (row=rstart; row<rend; row++) {
9876214f412SHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9886214f412SHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9896214f412SHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
9906214f412SHong Zhang     for (j=0; j<ncols; j++) { /* lower triangular part */
9916214f412SHong Zhang       if (cols[j] == row) continue;
9926214f412SHong Zhang       ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr);
9936214f412SHong Zhang     }
9946214f412SHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9956214f412SHong Zhang   }
9966214f412SHong Zhang   ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr);
9976214f412SHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9986214f412SHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9996214f412SHong Zhang 
10006214f412SHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
100128be2f97SBarry Smith     ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr);
10026214f412SHong Zhang   } else {
10036214f412SHong Zhang     *newmat = mat_elemental;
10046214f412SHong Zhang   }
10056214f412SHong Zhang   PetscFunctionReturn(0);
10066214f412SHong Zhang }
10076214f412SHong Zhang 
10086214f412SHong Zhang #undef __FUNCT__
1009db31f6deSJed Brown #define __FUNCT__ "MatDestroy_Elemental"
1010db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A)
1011db31f6deSJed Brown {
1012db31f6deSJed Brown   Mat_Elemental      *a = (Mat_Elemental*)A->data;
1013db31f6deSJed Brown   PetscErrorCode     ierr;
10145e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
10155e9f5b67SHong Zhang   PetscBool          flg;
10165e9f5b67SHong Zhang   MPI_Comm           icomm;
1017db31f6deSJed Brown 
1018db31f6deSJed Brown   PetscFunctionBegin;
1019c1ee1e62SHong Zhang   a->interface->Detach();
1020aae2c449SHong Zhang   delete a->interface;
1021aae2c449SHong Zhang   delete a->esubmat;
1022db31f6deSJed Brown   delete a->emat;
1023da0640a4SHong Zhang   delete a->pivot;
10245e9f5b67SHong Zhang 
1025e8146892SSatish Balay   El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
10260c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
10275e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
1028da0640a4SHong Zhang   /* printf("commgrid->grid_refct = %d, grid=%p\n",commgrid->grid_refct,commgrid->grid); -- memory leak revealed by valgrind? */
10295e9f5b67SHong Zhang   if (--commgrid->grid_refct == 0) {
10305e9f5b67SHong Zhang     delete commgrid->grid;
10315e9f5b67SHong Zhang     ierr = PetscFree(commgrid);CHKERRQ(ierr);
10325e9f5b67SHong Zhang   }
10335e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
1034bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr);
1035bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr);
1036d5560a0bSSatish Balay   ierr = PetscObjectComposeFunction((PetscObject)A,"MatElementalHermitianGenDefEig_C",NULL);CHKERRQ(ierr);
1037db31f6deSJed Brown   ierr = PetscFree(A->data);CHKERRQ(ierr);
1038db31f6deSJed Brown   PetscFunctionReturn(0);
1039db31f6deSJed Brown }
1040db31f6deSJed Brown 
1041db31f6deSJed Brown #undef __FUNCT__
1042db31f6deSJed Brown #define __FUNCT__ "MatSetUp_Elemental"
1043db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A)
1044db31f6deSJed Brown {
1045db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
1046db31f6deSJed Brown   PetscErrorCode ierr;
1047db31f6deSJed Brown   PetscMPIInt    rsize,csize;
1048db31f6deSJed Brown 
1049db31f6deSJed Brown   PetscFunctionBegin;
1050db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
1051db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
1052db31f6deSJed Brown 
1053efb79153SJack Poulson   a->emat->Resize(A->rmap->N,A->cmap->N);CHKERRQ(ierr);
1054e8146892SSatish Balay   El::Zero(*a->emat);
1055db31f6deSJed Brown 
1056db31f6deSJed Brown   ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr);
1057db31f6deSJed Brown   ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr);
1058ce94432eSBarry Smith   if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
1059db31f6deSJed Brown   a->commsize = rsize;
1060db31f6deSJed Brown   a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
1061db31f6deSJed Brown   a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
1062db31f6deSJed Brown   a->m[0]  = A->rmap->N / rsize + (a->mr[0] != rsize);
1063db31f6deSJed Brown   a->m[1]  = A->cmap->N / csize + (a->mr[1] != csize);
1064db31f6deSJed Brown   PetscFunctionReturn(0);
1065db31f6deSJed Brown }
1066db31f6deSJed Brown 
1067aae2c449SHong Zhang #undef __FUNCT__
1068aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyBegin_Elemental"
1069aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
1070aae2c449SHong Zhang {
1071aae2c449SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
1072aae2c449SHong Zhang 
1073aae2c449SHong Zhang   PetscFunctionBegin;
1074aae2c449SHong Zhang   a->interface->Detach();
1075e8146892SSatish Balay   a->interface->Attach(El::LOCAL_TO_GLOBAL,*(a->emat));
1076aae2c449SHong Zhang   PetscFunctionReturn(0);
1077aae2c449SHong Zhang }
1078aae2c449SHong Zhang 
1079aae2c449SHong Zhang #undef __FUNCT__
1080aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyEnd_Elemental"
1081aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
1082aae2c449SHong Zhang {
1083aae2c449SHong Zhang   PetscFunctionBegin;
1084aae2c449SHong Zhang   /* Currently does nothing */
1085aae2c449SHong Zhang   PetscFunctionReturn(0);
1086aae2c449SHong Zhang }
1087aae2c449SHong Zhang 
10887b30e0f1SHong Zhang #undef __FUNCT__
10897b30e0f1SHong Zhang #define __FUNCT__ "MatLoad_Elemental"
10907b30e0f1SHong Zhang PetscErrorCode MatLoad_Elemental(Mat newMat, PetscViewer viewer)
10917b30e0f1SHong Zhang {
10927b30e0f1SHong Zhang   PetscErrorCode ierr;
10937b30e0f1SHong Zhang   Mat            Adense,Ae;
10947b30e0f1SHong Zhang   MPI_Comm       comm;
10957b30e0f1SHong Zhang 
10967b30e0f1SHong Zhang   PetscFunctionBegin;
10977b30e0f1SHong Zhang   ierr = PetscObjectGetComm((PetscObject)newMat,&comm);CHKERRQ(ierr);
10987b30e0f1SHong Zhang   ierr = MatCreate(comm,&Adense);CHKERRQ(ierr);
10997b30e0f1SHong Zhang   ierr = MatSetType(Adense,MATDENSE);CHKERRQ(ierr);
11007b30e0f1SHong Zhang   ierr = MatLoad(Adense,viewer);CHKERRQ(ierr);
11017b30e0f1SHong Zhang   ierr = MatConvert(Adense, MATELEMENTAL, MAT_INITIAL_MATRIX,&Ae);CHKERRQ(ierr);
11027b30e0f1SHong Zhang   ierr = MatDestroy(&Adense);CHKERRQ(ierr);
110328be2f97SBarry Smith   ierr = MatHeaderReplace(newMat,&Ae);CHKERRQ(ierr);
11047b30e0f1SHong Zhang   PetscFunctionReturn(0);
11057b30e0f1SHong Zhang }
11067b30e0f1SHong Zhang 
11071d08bef3SHong Zhang #undef __FUNCT__
1108d5560a0bSSatish Balay #define __FUNCT__ "MatElementalHermitianGenDefEig_Elemental"
11092ef15aa2SHong Zhang PetscErrorCode MatElementalHermitianGenDefEig_Elemental(El::Pencil eigtype,El::UpperOrLower uplo,Mat A,Mat B,Mat *evals,Mat *evec,El::SortType sort,El::HermitianEigSubset<PetscElemScalar> subset,const El::HermitianEigCtrl<PetscElemScalar> ctrl)
11101d08bef3SHong Zhang {
11115dd12673SHong Zhang   PetscErrorCode ierr;
1112874e4880SHong Zhang   Mat_Elemental  *a=(Mat_Elemental*)A->data,*b=(Mat_Elemental*)B->data,*x;
1113874e4880SHong Zhang   MPI_Comm       comm;
1114874e4880SHong Zhang   Mat            EVAL;
11152ef15aa2SHong Zhang   Mat_Elemental  *e;
1116874e4880SHong Zhang 
11171d08bef3SHong Zhang   PetscFunctionBegin;
1118874e4880SHong Zhang   /* Compute eigenvalues and eigenvectors */
1119d5560a0bSSatish Balay   El::DistMatrix<PetscElemScalar,El::VR,El::STAR> w( *a->grid ); /* holding eigenvalues */
1120d5560a0bSSatish Balay   El::DistMatrix<PetscElemScalar>                 X( *a->grid ); /* holding eigenvectors */
1121935dd83bSHong Zhang   El::HermitianGenDefEig(eigtype,uplo,*a->emat,*b->emat,w,X,sort,subset,ctrl);
1122d5560a0bSSatish Balay   /* El::Print(w, "Eigenvalues"); */
11235dd12673SHong Zhang 
1124874e4880SHong Zhang   /* Wrap w and X into PETSc's MATMATELEMENTAL matrices */
1125d59fe0e1SHong Zhang   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
1126d59fe0e1SHong Zhang   ierr = MatCreate(comm,evec);CHKERRQ(ierr);
1127d59fe0e1SHong Zhang   ierr = MatSetSizes(*evec,PETSC_DECIDE,PETSC_DECIDE,X.Height(),X.Width());CHKERRQ(ierr);
1128d59fe0e1SHong Zhang   ierr = MatSetType(*evec,MATELEMENTAL);CHKERRQ(ierr);
1129d59fe0e1SHong Zhang   ierr = MatSetFromOptions(*evec);CHKERRQ(ierr);
1130d59fe0e1SHong Zhang   ierr = MatSetUp(*evec);CHKERRQ(ierr);
1131d59fe0e1SHong Zhang   ierr = MatAssemblyBegin(*evec,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1132d59fe0e1SHong Zhang   ierr = MatAssemblyEnd(*evec,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1133d59fe0e1SHong Zhang 
1134874e4880SHong Zhang   x = (Mat_Elemental*)(*evec)->data;
1135d59fe0e1SHong Zhang   //delete x->emat; //-- memory leak???
1136d59fe0e1SHong Zhang   *x->emat = X;
1137d59fe0e1SHong Zhang 
1138874e4880SHong Zhang   ierr = MatCreate(comm,&EVAL);CHKERRQ(ierr);
1139874e4880SHong Zhang   ierr = MatSetSizes(EVAL,PETSC_DECIDE,PETSC_DECIDE,w.Height(),w.Width());CHKERRQ(ierr);
1140874e4880SHong Zhang   ierr = MatSetType(EVAL,MATELEMENTAL);CHKERRQ(ierr);
1141874e4880SHong Zhang   ierr = MatSetFromOptions(EVAL);CHKERRQ(ierr);
1142874e4880SHong Zhang   ierr = MatSetUp(EVAL);CHKERRQ(ierr);
1143874e4880SHong Zhang   ierr = MatAssemblyBegin(EVAL,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1144874e4880SHong Zhang   ierr = MatAssemblyEnd(EVAL,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
11452ef15aa2SHong Zhang   e         = (Mat_Elemental*)EVAL->data;
1146874e4880SHong Zhang   *e->emat = w; //-- memory leak???
1147874e4880SHong Zhang   *evals   = EVAL;
1148874e4880SHong Zhang 
1149874e4880SHong Zhang #if defined(MV)
1150f82baa17SHong Zhang   /* Test correctness norm = || - A*X + B*X*w || */
1151d59fe0e1SHong Zhang   {
1152f82baa17SHong Zhang     PetscElemScalar alpha,beta;
1153d5560a0bSSatish Balay     El::DistMatrix<PetscElemScalar> Y(*a->grid); //tmp matrix
1154f82baa17SHong Zhang     alpha = 1.0; beta=0.0;
1155d5560a0bSSatish Balay     El::Gemm(El::NORMAL,El::NORMAL,alpha,*b->emat,X,beta,Y); //Y = B*X
1156d5560a0bSSatish Balay     El::DiagonalScale(El::RIGHT,El::NORMAL, w, Y); //Y = Y*w
1157f82baa17SHong Zhang     alpha = -1.0; beta=1.0;
1158d5560a0bSSatish Balay     El::Gemm(El::NORMAL,El::NORMAL,alpha,*a->emat,X,beta,Y); //Y = - A*X + B*X*w
1159f82baa17SHong Zhang 
1160d5560a0bSSatish Balay     PetscElemScalar norm = El::FrobeniusNorm(Y);
1161d59fe0e1SHong Zhang     if ((*a->grid).Rank()==0) printf("  norm (- A*X + B*X*w) = %g\n",norm);
1162d59fe0e1SHong Zhang   }
1163f82baa17SHong Zhang 
1164874e4880SHong Zhang   {
1165874e4880SHong Zhang     PetscMPIInt rank;
1166874e4880SHong Zhang     ierr = MPI_Comm_rank(comm,&rank);
1167874e4880SHong Zhang     printf("w: [%d] [%d, %d %d] %d; X: %d %d\n",rank,w.DistRank(),w.ColRank(),w.RowRank(),w.LocalHeight(),X.LocalHeight(),X.LocalWidth());
11685dd12673SHong Zhang   }
1169874e4880SHong Zhang #endif
11701d08bef3SHong Zhang   PetscFunctionReturn(0);
11711d08bef3SHong Zhang }
11721d08bef3SHong Zhang 
11731d08bef3SHong Zhang #undef __FUNCT__
1174d5560a0bSSatish Balay #define __FUNCT__ "MatElementalHermitianGenDefEig"
11751d08bef3SHong Zhang /*@
11762ef15aa2SHong Zhang   MatElementalHermitianGenDefEig - Compute the set of eigenvalues of the Hermitian-definite matrix pencil determined by the subset structure
11771d08bef3SHong Zhang 
11781d08bef3SHong Zhang    Logically Collective on Mat
11791d08bef3SHong Zhang 
11801d08bef3SHong Zhang    Level: beginner
11811d08bef3SHong Zhang 
11822ef15aa2SHong Zhang    References: Elemental Users' Guide
11831d08bef3SHong Zhang @*/
11842ef15aa2SHong Zhang PetscErrorCode MatElementalHermitianGenDefEig(El::Pencil type,El::UpperOrLower uplo,Mat A,Mat B,Mat *evals,Mat *evec,El::SortType sort,El::HermitianEigSubset<PetscElemScalar> subset,const El::HermitianEigCtrl<PetscElemScalar> ctrl)
11851d08bef3SHong Zhang {
11861d08bef3SHong Zhang   PetscErrorCode ierr;
11871d08bef3SHong Zhang 
11881d08bef3SHong Zhang   PetscFunctionBegin;
11892ef15aa2SHong Zhang   ierr = PetscTryMethod(A,"MatElementalHermitianGenDefEig_C",(El::Pencil,El::UpperOrLower,Mat,Mat,Mat*,Mat*,El::SortType,El::HermitianEigSubset<PetscElemScalar>,const El::HermitianEigCtrl<PetscElemScalar>),(type,uplo,A,B,evals,evec,sort,subset,ctrl));CHKERRQ(ierr);
11901d08bef3SHong Zhang   PetscFunctionReturn(0);
11911d08bef3SHong Zhang }
11921d08bef3SHong Zhang 
119340d92e34SHong Zhang /* -------------------------------------------------------------------*/
119440d92e34SHong Zhang static struct _MatOps MatOps_Values = {
119540d92e34SHong Zhang        MatSetValues_Elemental,
119640d92e34SHong Zhang        0,
119740d92e34SHong Zhang        0,
119840d92e34SHong Zhang        MatMult_Elemental,
119940d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental,
12009426833fSXuan Zhou        MatMultTranspose_Elemental,
1201e883f9d5SXuan Zhou        MatMultTransposeAdd_Elemental,
120240d92e34SHong Zhang        MatSolve_Elemental,
1203df311e6cSXuan Zhou        MatSolveAdd_Elemental,
12042ef15aa2SHong Zhang        0,
12052ef15aa2SHong Zhang /*10*/ 0,
120640d92e34SHong Zhang        MatLUFactor_Elemental,
120740d92e34SHong Zhang        MatCholeskyFactor_Elemental,
120840d92e34SHong Zhang        0,
120940d92e34SHong Zhang        MatTranspose_Elemental,
121040d92e34SHong Zhang /*15*/ MatGetInfo_Elemental,
121140d92e34SHong Zhang        0,
121261119200SXuan Zhou        MatGetDiagonal_Elemental,
1213ade3cc5eSXuan Zhou        MatDiagonalScale_Elemental,
121440d92e34SHong Zhang        MatNorm_Elemental,
121540d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental,
121640d92e34SHong Zhang        MatAssemblyEnd_Elemental,
12172ef15aa2SHong Zhang        0,
121840d92e34SHong Zhang        MatZeroEntries_Elemental,
121940d92e34SHong Zhang /*24*/ 0,
122040d92e34SHong Zhang        MatLUFactorSymbolic_Elemental,
122140d92e34SHong Zhang        MatLUFactorNumeric_Elemental,
122240d92e34SHong Zhang        MatCholeskyFactorSymbolic_Elemental,
122340d92e34SHong Zhang        MatCholeskyFactorNumeric_Elemental,
122440d92e34SHong Zhang /*29*/ MatSetUp_Elemental,
122540d92e34SHong Zhang        0,
122640d92e34SHong Zhang        0,
122740d92e34SHong Zhang        0,
122840d92e34SHong Zhang        0,
1229df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental,
123040d92e34SHong Zhang        0,
123140d92e34SHong Zhang        0,
123240d92e34SHong Zhang        0,
123340d92e34SHong Zhang        0,
123440d92e34SHong Zhang /*39*/ MatAXPY_Elemental,
123540d92e34SHong Zhang        0,
123640d92e34SHong Zhang        0,
123740d92e34SHong Zhang        0,
123840d92e34SHong Zhang        MatCopy_Elemental,
123940d92e34SHong Zhang /*44*/ 0,
124040d92e34SHong Zhang        MatScale_Elemental,
12417d68702bSBarry Smith        MatShift_Basic,
124240d92e34SHong Zhang        0,
124340d92e34SHong Zhang        0,
124440d92e34SHong Zhang /*49*/ 0,
124540d92e34SHong Zhang        0,
124640d92e34SHong Zhang        0,
124740d92e34SHong Zhang        0,
124840d92e34SHong Zhang        0,
124940d92e34SHong Zhang /*54*/ 0,
125040d92e34SHong Zhang        0,
125140d92e34SHong Zhang        0,
125240d92e34SHong Zhang        0,
125340d92e34SHong Zhang        0,
125440d92e34SHong Zhang /*59*/ 0,
125540d92e34SHong Zhang        MatDestroy_Elemental,
125640d92e34SHong Zhang        MatView_Elemental,
125740d92e34SHong Zhang        0,
125840d92e34SHong Zhang        0,
125940d92e34SHong Zhang /*64*/ 0,
126040d92e34SHong Zhang        0,
126140d92e34SHong Zhang        0,
126240d92e34SHong Zhang        0,
126340d92e34SHong Zhang        0,
126440d92e34SHong Zhang /*69*/ 0,
126540d92e34SHong Zhang        0,
12662ef0cf24SXuan Zhou        MatConvert_Elemental_Dense,
126740d92e34SHong Zhang        0,
126840d92e34SHong Zhang        0,
126940d92e34SHong Zhang /*74*/ 0,
127040d92e34SHong Zhang        0,
127140d92e34SHong Zhang        0,
127240d92e34SHong Zhang        0,
127340d92e34SHong Zhang        0,
127440d92e34SHong Zhang /*79*/ 0,
127540d92e34SHong Zhang        0,
127640d92e34SHong Zhang        0,
127740d92e34SHong Zhang        0,
12787b30e0f1SHong Zhang        MatLoad_Elemental,
127940d92e34SHong Zhang /*84*/ 0,
128040d92e34SHong Zhang        0,
128140d92e34SHong Zhang        0,
128240d92e34SHong Zhang        0,
128340d92e34SHong Zhang        0,
128440d92e34SHong Zhang /*89*/ MatMatMult_Elemental,
128540d92e34SHong Zhang        MatMatMultSymbolic_Elemental,
128640d92e34SHong Zhang        MatMatMultNumeric_Elemental,
128740d92e34SHong Zhang        0,
128840d92e34SHong Zhang        0,
128940d92e34SHong Zhang /*94*/ 0,
1290df311e6cSXuan Zhou        MatMatTransposeMult_Elemental,
1291df311e6cSXuan Zhou        MatMatTransposeMultSymbolic_Elemental,
1292df311e6cSXuan Zhou        MatMatTransposeMultNumeric_Elemental,
129340d92e34SHong Zhang        0,
129440d92e34SHong Zhang /*99*/ 0,
129540d92e34SHong Zhang        0,
129640d92e34SHong Zhang        0,
1297dfcb0403SXuan Zhou        MatConjugate_Elemental,
129840d92e34SHong Zhang        0,
129940d92e34SHong Zhang /*104*/0,
130040d92e34SHong Zhang        0,
130140d92e34SHong Zhang        0,
130240d92e34SHong Zhang        0,
130340d92e34SHong Zhang        0,
130440d92e34SHong Zhang /*109*/MatMatSolve_Elemental,
130540d92e34SHong Zhang        0,
130640d92e34SHong Zhang        0,
130740d92e34SHong Zhang        0,
130840d92e34SHong Zhang        0,
130940d92e34SHong Zhang /*114*/0,
131040d92e34SHong Zhang        0,
131140d92e34SHong Zhang        0,
131240d92e34SHong Zhang        0,
131340d92e34SHong Zhang        0,
131440d92e34SHong Zhang /*119*/0,
13154a29722dSXuan Zhou        MatHermitianTranspose_Elemental,
131640d92e34SHong Zhang        0,
131740d92e34SHong Zhang        0,
131840d92e34SHong Zhang        0,
131940d92e34SHong Zhang /*124*/0,
132040d92e34SHong Zhang        0,
132140d92e34SHong Zhang        0,
132240d92e34SHong Zhang        0,
132340d92e34SHong Zhang        0,
132440d92e34SHong Zhang /*129*/0,
132540d92e34SHong Zhang        0,
132640d92e34SHong Zhang        0,
132740d92e34SHong Zhang        0,
132840d92e34SHong Zhang        0,
132940d92e34SHong Zhang /*134*/0,
133040d92e34SHong Zhang        0,
133140d92e34SHong Zhang        0,
133240d92e34SHong Zhang        0,
133340d92e34SHong Zhang        0
133440d92e34SHong Zhang };
133540d92e34SHong Zhang 
1336ed36708cSHong Zhang /*MC
1337ed36708cSHong Zhang    MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1338ed36708cSHong Zhang 
1339c2b89b5dSBarry Smith   Use ./configure --download-elemental to install PETSc to use Elemental
1340c2b89b5dSBarry Smith 
1341c2b89b5dSBarry Smith   Use -pc_type lu -pc_factor_mat_solver_package elemental to us this direct solver
1342c2b89b5dSBarry Smith 
1343ed36708cSHong Zhang    Options Database Keys:
13445cc86fc1SJed Brown + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
13455cc86fc1SJed Brown - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1346ed36708cSHong Zhang 
1347ed36708cSHong Zhang   Level: beginner
1348ed36708cSHong Zhang 
13495cb544a0SHong Zhang .seealso: MATDENSE
1350ed36708cSHong Zhang M*/
13514a29722dSXuan Zhou 
1352db31f6deSJed Brown #undef __FUNCT__
1353db31f6deSJed Brown #define __FUNCT__ "MatCreate_Elemental"
13548cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1355db31f6deSJed Brown {
1356db31f6deSJed Brown   Mat_Elemental      *a;
1357db31f6deSJed Brown   PetscErrorCode     ierr;
13585682a260SJack Poulson   PetscBool          flg,flg1;
13595e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
13605e9f5b67SHong Zhang   MPI_Comm           icomm;
13615682a260SJack Poulson   PetscInt           optv1;
1362db31f6deSJed Brown 
1363db31f6deSJed Brown   PetscFunctionBegin;
1364607a6623SBarry Smith   ierr = PetscElementalInitializePackage();CHKERRQ(ierr);
136540d92e34SHong Zhang   ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
136640d92e34SHong Zhang   A->insertmode = NOT_SET_VALUES;
1367db31f6deSJed Brown 
1368b00a9115SJed Brown   ierr = PetscNewLog(A,&a);CHKERRQ(ierr);
1369db31f6deSJed Brown   A->data = (void*)a;
1370db31f6deSJed Brown 
1371db31f6deSJed Brown   /* Set up the elemental matrix */
1372e8146892SSatish Balay   El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
13735e9f5b67SHong Zhang 
13745e9f5b67SHong Zhang   /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
13755e9f5b67SHong Zhang   if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
1376180a43e4SHong Zhang     ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);
13775e9f5b67SHong Zhang   }
13780c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
13795e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
13805e9f5b67SHong Zhang   if (!flg) {
1381b00a9115SJed Brown     ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr);
13825cb544a0SHong Zhang 
1383ce94432eSBarry Smith     ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr);
1384e8146892SSatish Balay     /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until El::Grid() is called */
1385e8146892SSatish Balay     ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",El::mpi::Size(cxxcomm),&optv1,&flg1);CHKERRQ(ierr);
13865682a260SJack Poulson     if (flg1) {
1387e8146892SSatish Balay       if (El::mpi::Size(cxxcomm) % optv1 != 0) {
1388e8146892SSatish Balay         SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)El::mpi::Size(cxxcomm));
1389ed667823SXuan Zhou       }
1390e8146892SSatish Balay       commgrid->grid = new El::Grid(cxxcomm,optv1); /* use user-provided grid height */
13912ef0cf24SXuan Zhou     } else {
1392e8146892SSatish Balay       commgrid->grid = new El::Grid(cxxcomm); /* use Elemental default grid sizes */
1393da0640a4SHong Zhang       /* printf("new commgrid->grid = %p\n",commgrid->grid);  -- memory leak revealed by valgrind? */
1394ed667823SXuan Zhou     }
13955e9f5b67SHong Zhang     commgrid->grid_refct = 1;
13965e9f5b67SHong Zhang     ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr);
13975cb544a0SHong Zhang     PetscOptionsEnd();
13985e9f5b67SHong Zhang   } else {
13995e9f5b67SHong Zhang     commgrid->grid_refct++;
14005e9f5b67SHong Zhang   }
14015e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
14025e9f5b67SHong Zhang   a->grid      = commgrid->grid;
1403e8146892SSatish Balay   a->emat      = new El::DistMatrix<PetscElemScalar>(*a->grid);
1404e8146892SSatish Balay   a->esubmat   = new El::Matrix<PetscElemScalar>(1,1);
1405e8146892SSatish Balay   a->interface = new El::AxpyInterface<PetscElemScalar>;
1406e8146892SSatish Balay   a->pivot     = new El::DistMatrix<PetscInt,El::VC,El::STAR>;
1407db31f6deSJed Brown 
1408db31f6deSJed Brown   /* build cache for off array entries formed */
1409e8146892SSatish Balay   a->interface->Attach(El::LOCAL_TO_GLOBAL,*(a->emat));
1410bafd5131SHong Zhang 
1411bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr);
1412d5560a0bSSatish Balay   ierr = PetscObjectComposeFunction((PetscObject)A,"MatElementalHermitianGenDefEig_C",MatElementalHermitianGenDefEig_Elemental);CHKERRQ(ierr);
1413db31f6deSJed Brown 
1414db31f6deSJed Brown   ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr);
1415db31f6deSJed Brown   PetscFunctionReturn(0);
1416db31f6deSJed Brown }
1417