xref: /petsc/src/mat/impls/elemental/matelem.cxx (revision 382fd914c91bc298c4a8a26aba9727cdaa0dadbe)
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);
267b30e0f1SHong Zhang   //#if defined(MV)
27db31f6deSJed 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 */
28db31f6deSJed Brown     int zero = 0;
29db31f6deSJed Brown     char **nothing = 0;
30db31f6deSJed Brown     elem::Initialize(zero,nothing);
31db31f6deSJed Brown   }
327b30e0f1SHong Zhang   //#endif
33db31f6deSJed Brown   ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr);
34db31f6deSJed Brown   PetscFunctionReturn(0);
35db31f6deSJed Brown }
36db31f6deSJed Brown 
37db31f6deSJed Brown #undef __FUNCT__
38db31f6deSJed Brown #define __FUNCT__ "PetscElementalFinalizePackage"
39db31f6deSJed Brown /*@C
40db31f6deSJed Brown    PetscElementalFinalizePackage - Finalize Elemental package
41db31f6deSJed Brown 
42db31f6deSJed Brown    Logically Collective
43db31f6deSJed Brown 
44db31f6deSJed Brown    Level: developer
45db31f6deSJed Brown 
46db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage()
47db31f6deSJed Brown @*/
48db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void)
49db31f6deSJed Brown {
50db31f6deSJed Brown 
51db31f6deSJed Brown   PetscFunctionBegin;
52db31f6deSJed Brown   elem::Finalize();
53db31f6deSJed Brown   PetscFunctionReturn(0);
54db31f6deSJed Brown }
55db31f6deSJed Brown 
56db31f6deSJed Brown #undef __FUNCT__
57db31f6deSJed Brown #define __FUNCT__ "MatView_Elemental"
58db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer)
59db31f6deSJed Brown {
60db31f6deSJed Brown   PetscErrorCode ierr;
61db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
62db31f6deSJed Brown   PetscBool      iascii;
63db31f6deSJed Brown 
64db31f6deSJed Brown   PetscFunctionBegin;
65db31f6deSJed Brown   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
66db31f6deSJed Brown   if (iascii) {
67db31f6deSJed Brown     PetscViewerFormat format;
68db31f6deSJed Brown     ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
69db31f6deSJed Brown     if (format == PETSC_VIEWER_ASCII_INFO) {
7079673f7bSHong Zhang       /* call elemental viewing function */
712d8adcc7SHong Zhang       ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr);
72ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr);
73ed667823SXuan Zhou       ierr = PetscViewerASCIIPrintf(viewer,"  grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr);
744fe7bbcaSHong Zhang       if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
7579673f7bSHong Zhang         /* call elemental viewing function */
76ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr);
774fe7bbcaSHong Zhang       }
7879673f7bSHong Zhang 
79db31f6deSJed Brown     } else if (format == PETSC_VIEWER_DEFAULT) {
80db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr);
817a583510SJack Poulson       elem::Print( *a->emat, "Elemental matrix (cyclic ordering)" );
82db31f6deSJed Brown       ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr);
83834d3fecSHong Zhang       if (A->factortype == MAT_FACTOR_NONE){
8461119200SXuan Zhou         Mat Adense;
85ce94432eSBarry Smith         ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
8661119200SXuan Zhou         ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
8761119200SXuan Zhou         ierr = MatView(Adense,viewer);CHKERRQ(ierr);
8861119200SXuan Zhou         ierr = MatDestroy(&Adense);CHKERRQ(ierr);
89834d3fecSHong Zhang       }
90ce94432eSBarry Smith     } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format");
91d2daa67eSHong Zhang   } else {
925cb544a0SHong Zhang     /* convert to dense format and call MatView() */
9361119200SXuan Zhou     Mat Adense;
94ce94432eSBarry Smith     ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr);
9561119200SXuan Zhou     ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr);
9661119200SXuan Zhou     ierr = MatView(Adense,viewer);CHKERRQ(ierr);
9761119200SXuan Zhou     ierr = MatDestroy(&Adense);CHKERRQ(ierr);
98d2daa67eSHong Zhang   }
99db31f6deSJed Brown   PetscFunctionReturn(0);
100db31f6deSJed Brown }
101db31f6deSJed Brown 
102db31f6deSJed Brown #undef __FUNCT__
103180a43e4SHong Zhang #define __FUNCT__ "MatGetInfo_Elemental"
10415767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info)
105180a43e4SHong Zhang {
10615767789SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
10715767789SHong Zhang   PetscMPIInt    rank;
10815767789SHong Zhang 
109180a43e4SHong Zhang   PetscFunctionBegin;
110ce94432eSBarry Smith   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
11115767789SHong Zhang 
11215767789SHong Zhang   /* if (!rank) printf("          .........MatGetInfo_Elemental ...\n"); */
1135cb544a0SHong Zhang   info->block_size     = 1.0;
11415767789SHong Zhang 
11515767789SHong Zhang   if (flag == MAT_LOCAL) {
11615767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
11715767789SHong Zhang     info->nz_used        = info->nz_allocated;
11815767789SHong Zhang   } else if (flag == MAT_GLOBAL_MAX) {
119ce94432eSBarry Smith     //ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr);
12015767789SHong Zhang     /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */
12115767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet");
12215767789SHong Zhang   } else if (flag == MAT_GLOBAL_SUM) {
12315767789SHong Zhang     //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet");
12415767789SHong Zhang     info->nz_allocated   = (double)(*a->emat).AllocatedMemory(); /* locally allocated */
12515767789SHong Zhang     info->nz_used        = info->nz_allocated; /* assume Elemental does accurate allocation */
126ce94432eSBarry Smith     //ierr = MPI_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr);
12715767789SHong Zhang     //PetscPrintf(PETSC_COMM_SELF,"    ... [%d] locally allocated %g\n",rank,info->nz_allocated);
12815767789SHong Zhang   }
12915767789SHong Zhang 
13015767789SHong Zhang   info->nz_unneeded       = 0.0;
13115767789SHong Zhang   info->assemblies        = (double)A->num_ass;
13215767789SHong Zhang   info->mallocs           = 0;
13315767789SHong Zhang   info->memory            = ((PetscObject)A)->mem;
13415767789SHong Zhang   info->fill_ratio_given  = 0; /* determined by Elemental */
13515767789SHong Zhang   info->fill_ratio_needed = 0;
13615767789SHong Zhang   info->factor_mallocs    = 0;
137180a43e4SHong Zhang   PetscFunctionReturn(0);
138180a43e4SHong Zhang }
139180a43e4SHong Zhang 
140180a43e4SHong Zhang #undef __FUNCT__
141db31f6deSJed Brown #define __FUNCT__ "MatSetValues_Elemental"
142e6dea9dbSXuan Zhou static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode)
143db31f6deSJed Brown {
144db31f6deSJed Brown   PetscErrorCode ierr;
145db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
146db31f6deSJed Brown   PetscMPIInt    rank;
147db31f6deSJed Brown   PetscInt       i,j,rrank,ridx,crank,cidx;
148db31f6deSJed Brown 
149db31f6deSJed Brown   PetscFunctionBegin;
150ce94432eSBarry Smith   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr);
151db31f6deSJed Brown 
152db31f6deSJed Brown   const elem::Grid &grid = a->emat->Grid();
153db31f6deSJed Brown   for (i=0; i<nr; i++) {
154db31f6deSJed Brown     PetscInt erow,ecol,elrow,elcol;
155db31f6deSJed Brown     if (rows[i] < 0) continue;
156db31f6deSJed Brown     P2RO(A,0,rows[i],&rrank,&ridx);
157db31f6deSJed Brown     RO2E(A,0,rrank,ridx,&erow);
158ce94432eSBarry Smith     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation");
159db31f6deSJed Brown     for (j=0; j<nc; j++) {
160db31f6deSJed Brown       if (cols[j] < 0) continue;
161db31f6deSJed Brown       P2RO(A,1,cols[j],&crank,&cidx);
162db31f6deSJed Brown       RO2E(A,1,crank,cidx,&ecol);
163ce94432eSBarry Smith       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation");
164aae2c449SHong Zhang       if (erow % grid.MCSize() != grid.MCRank() || ecol % grid.MRSize() != grid.MRRank()){ /* off-proc entry */
165aae2c449SHong Zhang         if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported");
166aae2c449SHong 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); */
167e6dea9dbSXuan Zhou         a->esubmat->Set(0,0, (PetscElemScalar)vals[i*nc+j]);
168aae2c449SHong Zhang         a->interface->Axpy(1.0,*(a->esubmat),erow,ecol);
169aae2c449SHong Zhang         continue;
170ed36708cSHong Zhang       }
171db31f6deSJed Brown       elrow = erow / grid.MCSize();
172db31f6deSJed Brown       elcol = ecol / grid.MRSize();
173db31f6deSJed Brown       switch (imode) {
174e6dea9dbSXuan Zhou       case INSERT_VALUES: a->emat->SetLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
175e6dea9dbSXuan Zhou       case ADD_VALUES: a->emat->UpdateLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break;
176ce94432eSBarry Smith       default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode);
177db31f6deSJed Brown       }
178db31f6deSJed Brown     }
179db31f6deSJed Brown   }
180db31f6deSJed Brown   PetscFunctionReturn(0);
181db31f6deSJed Brown }
182db31f6deSJed Brown 
183db31f6deSJed Brown #undef __FUNCT__
184db31f6deSJed Brown #define __FUNCT__ "MatMult_Elemental"
185db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y)
186db31f6deSJed Brown {
187db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
188db31f6deSJed Brown   PetscErrorCode        ierr;
189e6dea9dbSXuan Zhou   const PetscElemScalar *x;
190e6dea9dbSXuan Zhou   PetscElemScalar       *y;
191df311e6cSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
192db31f6deSJed Brown 
193db31f6deSJed Brown   PetscFunctionBegin;
194e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
195e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
196db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
1970c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ye;
1980c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
1990c18141cSBarry Smith     ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n);
200db31f6deSJed Brown     elem::Gemv(elem::NORMAL,one,*a->emat,xe,zero,ye);
201db31f6deSJed Brown   }
202e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
203e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
204db31f6deSJed Brown   PetscFunctionReturn(0);
205db31f6deSJed Brown }
206db31f6deSJed Brown 
207db31f6deSJed Brown #undef __FUNCT__
2089426833fSXuan Zhou #define __FUNCT__ "MatMultTranspose_Elemental"
2099426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y)
2109426833fSXuan Zhou {
2119426833fSXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
2129426833fSXuan Zhou   PetscErrorCode        ierr;
213df311e6cSXuan Zhou   const PetscElemScalar *x;
214df311e6cSXuan Zhou   PetscElemScalar       *y;
215e6dea9dbSXuan Zhou   PetscElemScalar       one = 1,zero = 0;
2169426833fSXuan Zhou 
2179426833fSXuan Zhou   PetscFunctionBegin;
218e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
219e6dea9dbSXuan Zhou   ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2209426833fSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
2210c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ye;
2220c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
2230c18141cSBarry Smith     ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n);
2249426833fSXuan Zhou     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,zero,ye);
2259426833fSXuan Zhou   }
226e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
227e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr);
2289426833fSXuan Zhou   PetscFunctionReturn(0);
2299426833fSXuan Zhou }
2309426833fSXuan Zhou 
2319426833fSXuan Zhou #undef __FUNCT__
232db31f6deSJed Brown #define __FUNCT__ "MatMultAdd_Elemental"
233db31f6deSJed Brown static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
234db31f6deSJed Brown {
235db31f6deSJed Brown   Mat_Elemental         *a = (Mat_Elemental*)A->data;
236db31f6deSJed Brown   PetscErrorCode        ierr;
237df311e6cSXuan Zhou   const PetscElemScalar *x;
238df311e6cSXuan Zhou   PetscElemScalar       *z;
239e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
240db31f6deSJed Brown 
241db31f6deSJed Brown   PetscFunctionBegin;
242db31f6deSJed Brown   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
243e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
244e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
245db31f6deSJed Brown   { /* Scoping so that constructor is called before pointer is returned */
2460c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ze;
2470c18141cSBarry Smith     xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n);
2480c18141cSBarry Smith     ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n);
249db31f6deSJed Brown     elem::Gemv(elem::NORMAL,one,*a->emat,xe,one,ze);
250db31f6deSJed Brown   }
251e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
252e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
253db31f6deSJed Brown   PetscFunctionReturn(0);
254db31f6deSJed Brown }
255db31f6deSJed Brown 
256db31f6deSJed Brown #undef __FUNCT__
257e883f9d5SXuan Zhou #define __FUNCT__ "MatMultTransposeAdd_Elemental"
258e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z)
259e883f9d5SXuan Zhou {
260e883f9d5SXuan Zhou   Mat_Elemental         *a = (Mat_Elemental*)A->data;
261e883f9d5SXuan Zhou   PetscErrorCode        ierr;
262df311e6cSXuan Zhou   const PetscElemScalar *x;
263df311e6cSXuan Zhou   PetscElemScalar       *z;
264e6dea9dbSXuan Zhou   PetscElemScalar       one = 1;
265e883f9d5SXuan Zhou 
266e883f9d5SXuan Zhou   PetscFunctionBegin;
267e883f9d5SXuan Zhou   if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);}
268e6dea9dbSXuan Zhou   ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
269e6dea9dbSXuan Zhou   ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
270e883f9d5SXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
2710c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe, ze;
2720c18141cSBarry Smith     xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
2730c18141cSBarry Smith     ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n);
274e883f9d5SXuan Zhou     elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,one,ze);
275e883f9d5SXuan Zhou   }
276e6dea9dbSXuan Zhou   ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr);
277e6dea9dbSXuan Zhou   ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr);
278e883f9d5SXuan Zhou   PetscFunctionReturn(0);
279e883f9d5SXuan Zhou }
280e883f9d5SXuan Zhou 
281e883f9d5SXuan Zhou #undef __FUNCT__
2829a9e8502SHong Zhang #define __FUNCT__ "MatMatMultNumeric_Elemental"
2839a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C)
284c1d1b975SXuan Zhou {
285c1d1b975SXuan Zhou   Mat_Elemental    *a = (Mat_Elemental*)A->data;
286c1d1b975SXuan Zhou   Mat_Elemental    *b = (Mat_Elemental*)B->data;
2879a9e8502SHong Zhang   Mat_Elemental    *c = (Mat_Elemental*)C->data;
288e6dea9dbSXuan Zhou   PetscElemScalar  one = 1,zero = 0;
289c1d1b975SXuan Zhou 
290c1d1b975SXuan Zhou   PetscFunctionBegin;
291aae2c449SHong Zhang   { /* Scoping so that constructor is called before pointer is returned */
292c1d1b975SXuan Zhou     elem::Gemm(elem::NORMAL,elem::NORMAL,one,*a->emat,*b->emat,zero,*c->emat);
293aae2c449SHong Zhang   }
2949a9e8502SHong Zhang   C->assembled = PETSC_TRUE;
2959a9e8502SHong Zhang   PetscFunctionReturn(0);
2969a9e8502SHong Zhang }
2979a9e8502SHong Zhang 
2989a9e8502SHong Zhang #undef __FUNCT__
2999a9e8502SHong Zhang #define __FUNCT__ "MatMatMultSymbolic_Elemental"
3009a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
3019a9e8502SHong Zhang {
3029a9e8502SHong Zhang   PetscErrorCode ierr;
3039a9e8502SHong Zhang   Mat            Ce;
304ce94432eSBarry Smith   MPI_Comm       comm;
3059a9e8502SHong Zhang 
3069a9e8502SHong Zhang   PetscFunctionBegin;
307ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
3089a9e8502SHong Zhang   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
3099a9e8502SHong Zhang   ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
3109a9e8502SHong Zhang   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
3119a9e8502SHong Zhang   ierr = MatSetUp(Ce);CHKERRQ(ierr);
3129a9e8502SHong Zhang   *C = Ce;
3139a9e8502SHong Zhang   PetscFunctionReturn(0);
3149a9e8502SHong Zhang }
3159a9e8502SHong Zhang 
3169a9e8502SHong Zhang #undef __FUNCT__
3179a9e8502SHong Zhang #define __FUNCT__ "MatMatMult_Elemental"
3189a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3199a9e8502SHong Zhang {
3209a9e8502SHong Zhang   PetscErrorCode ierr;
3219a9e8502SHong Zhang 
3229a9e8502SHong Zhang   PetscFunctionBegin;
3239a9e8502SHong Zhang   if (scall == MAT_INITIAL_MATRIX){
3243ff4c91cSHong Zhang     ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3259a9e8502SHong Zhang     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
3263ff4c91cSHong Zhang     ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr);
3279a9e8502SHong Zhang   }
3283ff4c91cSHong Zhang   ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
3299a9e8502SHong Zhang   ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
3303ff4c91cSHong Zhang   ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr);
331c1d1b975SXuan Zhou   PetscFunctionReturn(0);
332c1d1b975SXuan Zhou }
333c1d1b975SXuan Zhou 
334c1d1b975SXuan Zhou #undef __FUNCT__
335df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultNumeric_Elemental"
336df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C)
337df311e6cSXuan Zhou {
338df311e6cSXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
339df311e6cSXuan Zhou   Mat_Elemental      *b = (Mat_Elemental*)B->data;
340df311e6cSXuan Zhou   Mat_Elemental      *c = (Mat_Elemental*)C->data;
341e6dea9dbSXuan Zhou   PetscElemScalar    one = 1,zero = 0;
342df311e6cSXuan Zhou 
343df311e6cSXuan Zhou   PetscFunctionBegin;
344df311e6cSXuan Zhou   { /* Scoping so that constructor is called before pointer is returned */
345df311e6cSXuan Zhou     elem::Gemm(elem::NORMAL,elem::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat);
346df311e6cSXuan Zhou   }
347df311e6cSXuan Zhou   C->assembled = PETSC_TRUE;
348df311e6cSXuan Zhou   PetscFunctionReturn(0);
349df311e6cSXuan Zhou }
350df311e6cSXuan Zhou 
351df311e6cSXuan Zhou #undef __FUNCT__
352df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultSymbolic_Elemental"
353df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C)
354df311e6cSXuan Zhou {
355df311e6cSXuan Zhou   PetscErrorCode ierr;
356df311e6cSXuan Zhou   Mat            Ce;
357ce94432eSBarry Smith   MPI_Comm       comm;
358df311e6cSXuan Zhou 
359df311e6cSXuan Zhou   PetscFunctionBegin;
360ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
361df311e6cSXuan Zhou   ierr = MatCreate(comm,&Ce);CHKERRQ(ierr);
362df311e6cSXuan Zhou   ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
363df311e6cSXuan Zhou   ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr);
364df311e6cSXuan Zhou   ierr = MatSetUp(Ce);CHKERRQ(ierr);
365df311e6cSXuan Zhou   *C = Ce;
366df311e6cSXuan Zhou   PetscFunctionReturn(0);
367df311e6cSXuan Zhou }
368df311e6cSXuan Zhou 
369df311e6cSXuan Zhou #undef __FUNCT__
370df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMult_Elemental"
371df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
372df311e6cSXuan Zhou {
373df311e6cSXuan Zhou   PetscErrorCode ierr;
374df311e6cSXuan Zhou 
375df311e6cSXuan Zhou   PetscFunctionBegin;
376df311e6cSXuan Zhou   if (scall == MAT_INITIAL_MATRIX){
377df311e6cSXuan Zhou     ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
378df311e6cSXuan Zhou     ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr);
379df311e6cSXuan Zhou     ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr);
380df311e6cSXuan Zhou   }
381df311e6cSXuan Zhou   ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
382df311e6cSXuan Zhou   ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr);
383df311e6cSXuan Zhou   ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr);
384df311e6cSXuan Zhou   PetscFunctionReturn(0);
385df311e6cSXuan Zhou }
386df311e6cSXuan Zhou 
387df311e6cSXuan Zhou #undef __FUNCT__
38861119200SXuan Zhou #define __FUNCT__ "MatGetDiagonal_Elemental"
389a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D)
39061119200SXuan Zhou {
391a9d89745SXuan Zhou   PetscInt        i,nrows,ncols,nD,rrank,ridx,crank,cidx;
392a9d89745SXuan Zhou   Mat_Elemental   *a = (Mat_Elemental*)A->data;
39361119200SXuan Zhou   PetscErrorCode  ierr;
394a9d89745SXuan Zhou   PetscElemScalar v;
395ce94432eSBarry Smith   MPI_Comm        comm;
39661119200SXuan Zhou 
39761119200SXuan Zhou   PetscFunctionBegin;
398ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
399a9d89745SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
400a9d89745SXuan Zhou   nD = nrows>ncols ? ncols : nrows;
401a9d89745SXuan Zhou   for (i=0; i<nD; i++) {
402a9d89745SXuan Zhou     PetscInt erow,ecol;
403a9d89745SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
404a9d89745SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
405a9d89745SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
406a9d89745SXuan Zhou     P2RO(A,1,i,&crank,&cidx);
407a9d89745SXuan Zhou     RO2E(A,1,crank,cidx,&ecol);
408a9d89745SXuan Zhou     if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
409a9d89745SXuan Zhou     v = a->emat->Get(erow,ecol);
4107c8b904fSSatish Balay     ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr);
411ade3cc5eSXuan Zhou   }
412a9d89745SXuan Zhou   ierr = VecAssemblyBegin(D);CHKERRQ(ierr);
413a9d89745SXuan Zhou   ierr = VecAssemblyEnd(D);CHKERRQ(ierr);
41461119200SXuan Zhou   PetscFunctionReturn(0);
41561119200SXuan Zhou }
41661119200SXuan Zhou 
41761119200SXuan Zhou #undef __FUNCT__
418ade3cc5eSXuan Zhou #define __FUNCT__ "MatDiagonalScale_Elemental"
419ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R)
420ade3cc5eSXuan Zhou {
421ade3cc5eSXuan Zhou   Mat_Elemental         *x = (Mat_Elemental*)X->data;
422ade3cc5eSXuan Zhou   const PetscElemScalar *d;
423ade3cc5eSXuan Zhou   PetscErrorCode        ierr;
424ade3cc5eSXuan Zhou 
425ade3cc5eSXuan Zhou   PetscFunctionBegin;
4269065cd98SJed Brown   if (R) {
427ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
4280c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de;
4290c18141cSBarry Smith     de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n);
430ade3cc5eSXuan Zhou     elem::DiagonalScale(elem::RIGHT,elem::NORMAL,de,*x->emat);
431ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr);
4329065cd98SJed Brown   }
4339065cd98SJed Brown   if (L) {
434ade3cc5eSXuan Zhou     ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
4350c18141cSBarry Smith     elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de;
4360c18141cSBarry Smith     de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n);
437ade3cc5eSXuan Zhou     elem::DiagonalScale(elem::LEFT,elem::NORMAL,de,*x->emat);
438ade3cc5eSXuan Zhou     ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr);
439ade3cc5eSXuan Zhou   }
440ade3cc5eSXuan Zhou   PetscFunctionReturn(0);
441ade3cc5eSXuan Zhou }
442ade3cc5eSXuan Zhou 
443ade3cc5eSXuan Zhou #undef __FUNCT__
4444ee44ac6SXuan Zhou #define __FUNCT__ "MatScale_Elemental"
445e6dea9dbSXuan Zhou static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a)
44665b78793SXuan Zhou {
44765b78793SXuan Zhou   Mat_Elemental  *x = (Mat_Elemental*)X->data;
44865b78793SXuan Zhou 
44965b78793SXuan Zhou   PetscFunctionBegin;
4500c18141cSBarry Smith   elem::Scale((PetscElemScalar)a,*x->emat);
45165b78793SXuan Zhou   PetscFunctionReturn(0);
45265b78793SXuan Zhou }
45365b78793SXuan Zhou 
454f82baa17SHong Zhang /*
455f82baa17SHong Zhang   MatAXPY - Computes Y = a*X + Y.
456f82baa17SHong Zhang */
45765b78793SXuan Zhou #undef __FUNCT__
4584ee44ac6SXuan Zhou #define __FUNCT__ "MatAXPY_Elemental"
459e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str)
460e09a3074SHong Zhang {
461e09a3074SHong Zhang   Mat_Elemental  *x = (Mat_Elemental*)X->data;
462e09a3074SHong Zhang   Mat_Elemental  *y = (Mat_Elemental*)Y->data;
46368446cf8SJed Brown   PetscErrorCode ierr;
464e09a3074SHong Zhang 
465e09a3074SHong Zhang   PetscFunctionBegin;
466e6dea9dbSXuan Zhou   elem::Axpy((PetscElemScalar)a,*x->emat,*y->emat);
46721f6c9c4SJed Brown   ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr);
468e09a3074SHong Zhang   PetscFunctionReturn(0);
469e09a3074SHong Zhang }
470e09a3074SHong Zhang 
471ae844d54SHong Zhang #undef __FUNCT__
472d6223691SXuan Zhou #define __FUNCT__ "MatCopy_Elemental"
473d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str)
474d6223691SXuan Zhou {
475d6223691SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
476d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
477d6223691SXuan Zhou 
478d6223691SXuan Zhou   PetscFunctionBegin;
479d6223691SXuan Zhou   elem::Copy(*a->emat,*b->emat);
480d6223691SXuan Zhou   PetscFunctionReturn(0);
481d6223691SXuan Zhou }
482d6223691SXuan Zhou 
483d6223691SXuan Zhou #undef __FUNCT__
484df311e6cSXuan Zhou #define __FUNCT__ "MatDuplicate_Elemental"
485df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B)
486df311e6cSXuan Zhou {
487df311e6cSXuan Zhou   Mat            Be;
488ce94432eSBarry Smith   MPI_Comm       comm;
489df311e6cSXuan Zhou   Mat_Elemental  *a=(Mat_Elemental*)A->data;
490df311e6cSXuan Zhou   PetscErrorCode ierr;
491df311e6cSXuan Zhou 
492df311e6cSXuan Zhou   PetscFunctionBegin;
493ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
494df311e6cSXuan Zhou   ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
495df311e6cSXuan Zhou   ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
496df311e6cSXuan Zhou   ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
497df311e6cSXuan Zhou   ierr = MatSetUp(Be);CHKERRQ(ierr);
498df311e6cSXuan Zhou   *B = Be;
499df311e6cSXuan Zhou   if (op == MAT_COPY_VALUES) {
500df311e6cSXuan Zhou     Mat_Elemental *b=(Mat_Elemental*)Be->data;
501df311e6cSXuan Zhou     elem::Copy(*a->emat,*b->emat);
502df311e6cSXuan Zhou   }
503df311e6cSXuan Zhou   Be->assembled = PETSC_TRUE;
504df311e6cSXuan Zhou   PetscFunctionReturn(0);
505df311e6cSXuan Zhou }
506df311e6cSXuan Zhou 
507df311e6cSXuan Zhou #undef __FUNCT__
508d6223691SXuan Zhou #define __FUNCT__ "MatTranspose_Elemental"
509d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
510d6223691SXuan Zhou {
511ec8cb81fSBarry Smith   Mat            Be = *B;
5125262d616SXuan Zhou   PetscErrorCode ierr;
513ce94432eSBarry Smith   MPI_Comm       comm;
5144fe7bbcaSHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
515d6223691SXuan Zhou 
516d6223691SXuan Zhou   PetscFunctionBegin;
517ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5185cb544a0SHong Zhang   /* Only out-of-place supported */
5195262d616SXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5205262d616SXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5215262d616SXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5225262d616SXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5235262d616SXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5245262d616SXuan Zhou     *B = Be;
5255262d616SXuan Zhou   }
5264fe7bbcaSHong Zhang   b = (Mat_Elemental*)Be->data;
5275262d616SXuan Zhou   elem::Transpose(*a->emat,*b->emat);
5285262d616SXuan Zhou   Be->assembled = PETSC_TRUE;
529d6223691SXuan Zhou   PetscFunctionReturn(0);
530d6223691SXuan Zhou }
531d6223691SXuan Zhou 
532d6223691SXuan Zhou #undef __FUNCT__
533dfcb0403SXuan Zhou #define __FUNCT__ "MatConjugate_Elemental"
534dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A)
535dfcb0403SXuan Zhou {
536dfcb0403SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
537dfcb0403SXuan Zhou 
538dfcb0403SXuan Zhou   PetscFunctionBegin;
539dfcb0403SXuan Zhou   elem::Conjugate(*a->emat);
540dfcb0403SXuan Zhou   PetscFunctionReturn(0);
541dfcb0403SXuan Zhou }
542dfcb0403SXuan Zhou 
543dfcb0403SXuan Zhou #undef __FUNCT__
5444a29722dSXuan Zhou #define __FUNCT__ "MatHermitianTranspose_Elemental"
5454a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B)
5464a29722dSXuan Zhou {
547ec8cb81fSBarry Smith   Mat            Be = *B;
5484a29722dSXuan Zhou   PetscErrorCode ierr;
549ce94432eSBarry Smith   MPI_Comm       comm;
5504a29722dSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data, *b;
5514a29722dSXuan Zhou 
5524a29722dSXuan Zhou   PetscFunctionBegin;
553ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
5545cb544a0SHong Zhang   /* Only out-of-place supported */
5554a29722dSXuan Zhou   if (reuse == MAT_INITIAL_MATRIX){
5564a29722dSXuan Zhou     ierr = MatCreate(comm,&Be);CHKERRQ(ierr);
5574a29722dSXuan Zhou     ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
5584a29722dSXuan Zhou     ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr);
5594a29722dSXuan Zhou     ierr = MatSetUp(Be);CHKERRQ(ierr);
5604a29722dSXuan Zhou     *B = Be;
5614a29722dSXuan Zhou   }
5624a29722dSXuan Zhou   b = (Mat_Elemental*)Be->data;
5634a29722dSXuan Zhou   elem::Adjoint(*a->emat,*b->emat);
5644a29722dSXuan Zhou   Be->assembled = PETSC_TRUE;
5654a29722dSXuan Zhou   PetscFunctionReturn(0);
5664a29722dSXuan Zhou }
5674a29722dSXuan Zhou 
5684a29722dSXuan Zhou #undef __FUNCT__
5691f881ff8SXuan Zhou #define __FUNCT__ "MatSolve_Elemental"
5701f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X)
5711f881ff8SXuan Zhou {
5721f881ff8SXuan Zhou   Mat_Elemental     *a = (Mat_Elemental*)A->data;
5731f881ff8SXuan Zhou   PetscErrorCode    ierr;
574df311e6cSXuan Zhou   PetscElemScalar   *x;
5751f881ff8SXuan Zhou 
5761f881ff8SXuan Zhou   PetscFunctionBegin;
57745cf121fSXuan Zhou   ierr = VecCopy(B,X);CHKERRQ(ierr);
578e6dea9dbSXuan Zhou   ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
5790c18141cSBarry Smith   elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe;
5800c18141cSBarry Smith   xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n);
5810c18141cSBarry Smith   elem::DistMatrix<PetscElemScalar,elem::MC,elem::MR> xer(xe);
582fc54b460SXuan Zhou   switch (A->factortype) {
583fc54b460SXuan Zhou   case MAT_FACTOR_LU:
5841f881ff8SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
585324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,xer);
58645cf121fSXuan Zhou       elem::Copy(xer,xe);
587fc54b460SXuan Zhou     } else {
588324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,xer);
58945cf121fSXuan Zhou       elem::Copy(xer,xe);
5901f881ff8SXuan Zhou     }
591fc54b460SXuan Zhou     break;
592fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
593324645c7SJack Poulson     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,xer);
594fc54b460SXuan Zhou     elem::Copy(xer,xe);
595fc54b460SXuan Zhou     break;
596fc54b460SXuan Zhou   default:
5974fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
598fc54b460SXuan Zhou     break;
5991f881ff8SXuan Zhou   }
600e6dea9dbSXuan Zhou   ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr);
6011f881ff8SXuan Zhou   PetscFunctionReturn(0);
6021f881ff8SXuan Zhou }
6031f881ff8SXuan Zhou 
6041f881ff8SXuan Zhou #undef __FUNCT__
605df311e6cSXuan Zhou #define __FUNCT__ "MatSolveAdd_Elemental"
606df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X)
607df311e6cSXuan Zhou {
608df311e6cSXuan Zhou   PetscErrorCode    ierr;
609df311e6cSXuan Zhou 
610df311e6cSXuan Zhou   PetscFunctionBegin;
611df311e6cSXuan Zhou   ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr);
6123d7f40dbSXuan Zhou   ierr = VecAXPY(X,1,Y);CHKERRQ(ierr);
613df311e6cSXuan Zhou   PetscFunctionReturn(0);
614df311e6cSXuan Zhou }
615df311e6cSXuan Zhou 
616df311e6cSXuan Zhou #undef __FUNCT__
617ae844d54SHong Zhang #define __FUNCT__ "MatMatSolve_Elemental"
618ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X)
619ae844d54SHong Zhang {
6201f0e42cfSHong Zhang   Mat_Elemental *a=(Mat_Elemental*)A->data;
621d6223691SXuan Zhou   Mat_Elemental *b=(Mat_Elemental*)B->data;
6221f0e42cfSHong Zhang   Mat_Elemental *x=(Mat_Elemental*)X->data;
6231f0e42cfSHong Zhang 
624ae844d54SHong Zhang   PetscFunctionBegin;
625d6223691SXuan Zhou   elem::Copy(*b->emat,*x->emat);
626fc54b460SXuan Zhou   switch (A->factortype) {
627fc54b460SXuan Zhou   case MAT_FACTOR_LU:
628d6223691SXuan Zhou     if ((*a->pivot).AllocatedMemory()) {
629324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*a->pivot,*x->emat);
630fc54b460SXuan Zhou     } else {
631324645c7SJack Poulson       elem::lu::SolveAfter(elem::NORMAL,*a->emat,*x->emat);
632d6223691SXuan Zhou     }
633fc54b460SXuan Zhou     break;
634fc54b460SXuan Zhou   case MAT_FACTOR_CHOLESKY:
635324645c7SJack Poulson     elem::cholesky::SolveAfter(elem::UPPER,elem::NORMAL,*a->emat,*x->emat);
636fc54b460SXuan Zhou     break;
637fc54b460SXuan Zhou   default:
6384fe7bbcaSHong Zhang     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType");
639fc54b460SXuan Zhou     break;
640fc54b460SXuan Zhou   }
641ae844d54SHong Zhang   PetscFunctionReturn(0);
642ae844d54SHong Zhang }
643ae844d54SHong Zhang 
644ae844d54SHong Zhang #undef __FUNCT__
645ae844d54SHong Zhang #define __FUNCT__ "MatLUFactor_Elemental"
646ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info)
647ae844d54SHong Zhang {
6487c920d81SXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
6497c920d81SXuan Zhou 
650ae844d54SHong Zhang   PetscFunctionBegin;
651d6223691SXuan Zhou   if (info->dtcol){
6527c920d81SXuan Zhou     elem::LU(*a->emat,*a->pivot);
6531293973dSHong Zhang   } else {
654d6223691SXuan Zhou     elem::LU(*a->emat);
655d6223691SXuan Zhou   }
6561293973dSHong Zhang   A->factortype = MAT_FACTOR_LU;
657834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
658ae844d54SHong Zhang   PetscFunctionReturn(0);
659ae844d54SHong Zhang }
660ae844d54SHong Zhang 
661d7c3f9d8SHong Zhang #undef __FUNCT__
662d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorNumeric_Elemental"
663d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
664d7c3f9d8SHong Zhang {
665d7c3f9d8SHong Zhang   PetscErrorCode ierr;
666d7c3f9d8SHong Zhang 
667d7c3f9d8SHong Zhang   PetscFunctionBegin;
668d7c3f9d8SHong Zhang   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
669d7c3f9d8SHong Zhang   ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr);
670d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
671d7c3f9d8SHong Zhang }
672d7c3f9d8SHong Zhang 
673d7c3f9d8SHong Zhang #undef __FUNCT__
674d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorSymbolic_Elemental"
675d7c3f9d8SHong Zhang static PetscErrorCode  MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
676d7c3f9d8SHong Zhang {
677d7c3f9d8SHong Zhang   PetscFunctionBegin;
678d7c3f9d8SHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
679d7c3f9d8SHong Zhang   PetscFunctionReturn(0);
680d7c3f9d8SHong Zhang }
681d7c3f9d8SHong Zhang 
68245cf121fSXuan Zhou #undef __FUNCT__
68345cf121fSXuan Zhou #define __FUNCT__ "MatCholeskyFactor_Elemental"
68445cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info)
68545cf121fSXuan Zhou {
68645cf121fSXuan Zhou   Mat_Elemental  *a = (Mat_Elemental*)A->data;
687df311e6cSXuan Zhou   elem::DistMatrix<PetscElemScalar,elem::MC,elem::STAR> d;
68845cf121fSXuan Zhou 
68945cf121fSXuan Zhou   PetscFunctionBegin;
690c9fc186eSXuan Zhou   elem::Cholesky(elem::UPPER,*a->emat);
691fc54b460SXuan Zhou   A->factortype = MAT_FACTOR_CHOLESKY;
692834d3fecSHong Zhang   A->assembled  = PETSC_TRUE;
69345cf121fSXuan Zhou   PetscFunctionReturn(0);
69445cf121fSXuan Zhou }
69545cf121fSXuan Zhou 
69679673f7bSHong Zhang #undef __FUNCT__
69779673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorNumeric_Elemental"
69879673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info)
69979673f7bSHong Zhang {
700cb76c1d8SXuan Zhou   PetscErrorCode ierr;
701cb76c1d8SXuan Zhou 
702cb76c1d8SXuan Zhou   PetscFunctionBegin;
703cb76c1d8SXuan Zhou   ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
704cb76c1d8SXuan Zhou   ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr);
705cb76c1d8SXuan Zhou   PetscFunctionReturn(0);
70679673f7bSHong Zhang }
707cb76c1d8SXuan Zhou 
70879673f7bSHong Zhang #undef __FUNCT__
70979673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorSymbolic_Elemental"
71079673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info)
71179673f7bSHong Zhang {
71279673f7bSHong Zhang   PetscFunctionBegin;
71379673f7bSHong Zhang   /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */
71479673f7bSHong Zhang   PetscFunctionReturn(0);
71579673f7bSHong Zhang }
71679673f7bSHong Zhang 
7171293973dSHong Zhang #undef __FUNCT__
71815767789SHong Zhang #define __FUNCT__ "MatFactorGetSolverPackage_elemental_elemental"
71915767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type)
72015767789SHong Zhang {
72115767789SHong Zhang   PetscFunctionBegin;
72215767789SHong Zhang   *type = MATSOLVERELEMENTAL;
72315767789SHong Zhang   PetscFunctionReturn(0);
72415767789SHong Zhang }
72515767789SHong Zhang 
72615767789SHong Zhang #undef __FUNCT__
72715767789SHong Zhang #define __FUNCT__ "MatGetFactor_elemental_elemental"
72815767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F)
7291293973dSHong Zhang {
7301293973dSHong Zhang   Mat            B;
7311293973dSHong Zhang   PetscErrorCode ierr;
7321293973dSHong Zhang 
7331293973dSHong Zhang   PetscFunctionBegin;
7341293973dSHong Zhang   /* Create the factorization matrix */
735ce94432eSBarry Smith   ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
7361293973dSHong Zhang   ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
7371293973dSHong Zhang   ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr);
7381293973dSHong Zhang   ierr = MatSetUp(B);CHKERRQ(ierr);
7391293973dSHong Zhang   B->factortype = ftype;
740bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr);
7411293973dSHong Zhang   *F            = B;
7421293973dSHong Zhang   PetscFunctionReturn(0);
7431293973dSHong Zhang }
7441293973dSHong Zhang 
7451f881ff8SXuan Zhou #undef __FUNCT__
74642c9c57cSBarry Smith #define __FUNCT__ "MatSolverPackageRegister_Elemental"
74729b38603SBarry Smith PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_Elemental(void)
74842c9c57cSBarry Smith {
74918713533SBarry Smith   PetscErrorCode ierr;
75018713533SBarry Smith 
75142c9c57cSBarry Smith   PetscFunctionBegin;
75242c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_LU,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
75342c9c57cSBarry Smith   ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL,        MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);CHKERRQ(ierr);
75442c9c57cSBarry Smith   PetscFunctionReturn(0);
75542c9c57cSBarry Smith }
75642c9c57cSBarry Smith 
75742c9c57cSBarry Smith #undef __FUNCT__
7581f881ff8SXuan Zhou #define __FUNCT__ "MatNorm_Elemental"
7591f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm)
7601f881ff8SXuan Zhou {
7611f881ff8SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7621f881ff8SXuan Zhou 
7631f881ff8SXuan Zhou   PetscFunctionBegin;
7641f881ff8SXuan Zhou   switch (type){
7651f881ff8SXuan Zhou   case NORM_1:
766324645c7SJack Poulson     *nrm = elem::OneNorm(*a->emat);
7671f881ff8SXuan Zhou     break;
7681f881ff8SXuan Zhou   case NORM_FROBENIUS:
769324645c7SJack Poulson     *nrm = elem::FrobeniusNorm(*a->emat);
7701f881ff8SXuan Zhou     break;
7711f881ff8SXuan Zhou   case NORM_INFINITY:
772324645c7SJack Poulson     *nrm = elem::InfinityNorm(*a->emat);
7731f881ff8SXuan Zhou     break;
7741f881ff8SXuan Zhou   default:
7751f881ff8SXuan Zhou     printf("Error: unsupported norm type!\n");
7761f881ff8SXuan Zhou   }
7771f881ff8SXuan Zhou   PetscFunctionReturn(0);
7781f881ff8SXuan Zhou }
7791f881ff8SXuan Zhou 
7805262d616SXuan Zhou #undef __FUNCT__
7815262d616SXuan Zhou #define __FUNCT__ "MatZeroEntries_Elemental"
7825262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A)
7835262d616SXuan Zhou {
7845262d616SXuan Zhou   Mat_Elemental *a=(Mat_Elemental*)A->data;
7855262d616SXuan Zhou 
7865262d616SXuan Zhou   PetscFunctionBegin;
7875262d616SXuan Zhou   elem::Zero(*a->emat);
7885262d616SXuan Zhou   PetscFunctionReturn(0);
7895262d616SXuan Zhou }
7905262d616SXuan Zhou 
791e09a3074SHong Zhang #undef __FUNCT__
792db31f6deSJed Brown #define __FUNCT__ "MatGetOwnershipIS_Elemental"
793db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols)
794db31f6deSJed Brown {
795db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
796db31f6deSJed Brown   PetscErrorCode ierr;
797db31f6deSJed Brown   PetscInt       i,m,shift,stride,*idx;
798db31f6deSJed Brown 
799db31f6deSJed Brown   PetscFunctionBegin;
800db31f6deSJed Brown   if (rows) {
801db31f6deSJed Brown     m = a->emat->LocalHeight();
802db31f6deSJed Brown     shift = a->emat->ColShift();
803db31f6deSJed Brown     stride = a->emat->ColStride();
804785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
805db31f6deSJed Brown     for (i=0; i<m; i++) {
806db31f6deSJed Brown       PetscInt rank,offset;
807db31f6deSJed Brown       E2RO(A,0,shift+i*stride,&rank,&offset);
808db31f6deSJed Brown       RO2P(A,0,rank,offset,&idx[i]);
809db31f6deSJed Brown     }
810db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
811db31f6deSJed Brown   }
812db31f6deSJed Brown   if (cols) {
813db31f6deSJed Brown     m = a->emat->LocalWidth();
814db31f6deSJed Brown     shift = a->emat->RowShift();
815db31f6deSJed Brown     stride = a->emat->RowStride();
816785e854fSJed Brown     ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr);
817db31f6deSJed Brown     for (i=0; i<m; i++) {
818db31f6deSJed Brown       PetscInt rank,offset;
819db31f6deSJed Brown       E2RO(A,1,shift+i*stride,&rank,&offset);
820db31f6deSJed Brown       RO2P(A,1,rank,offset,&idx[i]);
821db31f6deSJed Brown     }
822db31f6deSJed Brown     ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr);
823db31f6deSJed Brown   }
824db31f6deSJed Brown   PetscFunctionReturn(0);
825db31f6deSJed Brown }
826db31f6deSJed Brown 
827db31f6deSJed Brown #undef __FUNCT__
8282ef0cf24SXuan Zhou #define __FUNCT__ "MatConvert_Elemental_Dense"
82919fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B)
830af295397SXuan Zhou {
8312ef0cf24SXuan Zhou   Mat                Bmpi;
832af295397SXuan Zhou   Mat_Elemental      *a = (Mat_Elemental*)A->data;
833ce94432eSBarry Smith   MPI_Comm           comm;
8342ef0cf24SXuan Zhou   PetscErrorCode     ierr;
8352ef0cf24SXuan Zhou   PetscInt           rrank,ridx,crank,cidx,nrows,ncols,i,j;
836df311e6cSXuan Zhou   PetscElemScalar    v;
837573b0fb4SBarry Smith   PetscBool          s1,s2,s3;
838af295397SXuan Zhou 
839af295397SXuan Zhou   PetscFunctionBegin;
840ce94432eSBarry Smith   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
841573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATDENSE,&s1);CHKERRQ(ierr);
842573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATSEQDENSE,&s2);CHKERRQ(ierr);
843573b0fb4SBarry Smith   ierr = PetscStrcmp(newtype,MATMPIDENSE,&s3);CHKERRQ(ierr);
844573b0fb4SBarry Smith   if (!s1 && !s2 && !s3) SETERRQ(comm,PETSC_ERR_SUP,"Unsupported New MatType: must be MATDENSE, MATSEQDENSE or MATMPIDENSE");
845af295397SXuan Zhou   ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr);
846af295397SXuan Zhou   ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
8472ef0cf24SXuan Zhou   ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr);
848af295397SXuan Zhou   ierr = MatSetUp(Bmpi);CHKERRQ(ierr);
8492ef0cf24SXuan Zhou   ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr);
8502ef0cf24SXuan Zhou   for (i=0; i<nrows; i++) {
8512ef0cf24SXuan Zhou     PetscInt erow,ecol;
8522ef0cf24SXuan Zhou     P2RO(A,0,i,&rrank,&ridx);
8532ef0cf24SXuan Zhou     RO2E(A,0,rrank,ridx,&erow);
8542ef0cf24SXuan Zhou     if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation");
8552ef0cf24SXuan Zhou     for (j=0; j<ncols; j++) {
8562ef0cf24SXuan Zhou       P2RO(A,1,j,&crank,&cidx);
8572ef0cf24SXuan Zhou       RO2E(A,1,crank,cidx,&ecol);
8582ef0cf24SXuan Zhou       if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation");
8592ef0cf24SXuan Zhou       v = a->emat->Get(erow,ecol);
860e6dea9dbSXuan Zhou       ierr = MatSetValues(Bmpi,1,&i,1,&j,(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr);
8612ef0cf24SXuan Zhou     }
8622ef0cf24SXuan Zhou   }
863af295397SXuan Zhou   ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
864af295397SXuan Zhou   ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
865c4ad791aSXuan Zhou   if (reuse == MAT_REUSE_MATRIX) {
866c4ad791aSXuan Zhou     ierr = MatHeaderReplace(A,Bmpi);CHKERRQ(ierr);
867c4ad791aSXuan Zhou   } else {
868c4ad791aSXuan Zhou     *B = Bmpi;
869c4ad791aSXuan Zhou   }
870af295397SXuan Zhou   PetscFunctionReturn(0);
871af295397SXuan Zhou }
872af295397SXuan Zhou 
873af295397SXuan Zhou #undef __FUNCT__
874af8000cdSHong Zhang #define __FUNCT__ "MatConvert_SeqAIJ_Elemental"
875af8000cdSHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
876af8000cdSHong Zhang {
877af8000cdSHong Zhang   Mat               mat_elemental;
878af8000cdSHong Zhang   PetscErrorCode    ierr;
879af8000cdSHong Zhang   PetscInt          M=A->rmap->N,N=A->cmap->N,row,ncols;
880af8000cdSHong Zhang   const PetscInt    *cols;
881af8000cdSHong Zhang   const PetscScalar *vals;
882af8000cdSHong Zhang 
883af8000cdSHong Zhang   PetscFunctionBegin;
884af8000cdSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
885af8000cdSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
886af8000cdSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
887af8000cdSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
888af8000cdSHong Zhang   for (row=0; row<M; row++) {
889af8000cdSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
890af8000cdSHong Zhang     /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
891af8000cdSHong Zhang     ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr);
892af8000cdSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
893af8000cdSHong Zhang   }
894af8000cdSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
895af8000cdSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
896af8000cdSHong Zhang 
897af8000cdSHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
898af8000cdSHong Zhang     ierr = MatHeaderReplace(A,mat_elemental);CHKERRQ(ierr);
899af8000cdSHong Zhang   } else {
900af8000cdSHong Zhang     *newmat = mat_elemental;
901af8000cdSHong Zhang   }
902af8000cdSHong Zhang   PetscFunctionReturn(0);
903af8000cdSHong Zhang }
904af8000cdSHong Zhang 
905af8000cdSHong Zhang #undef __FUNCT__
9065d7652ecSHong Zhang #define __FUNCT__ "MatConvert_MPIAIJ_Elemental"
9075d7652ecSHong Zhang PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat)
9085d7652ecSHong Zhang {
9095d7652ecSHong Zhang   Mat               mat_elemental;
9105d7652ecSHong Zhang   PetscErrorCode    ierr;
9115d7652ecSHong Zhang   PetscInt          row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j;
9125d7652ecSHong Zhang   const PetscInt    *cols;
9135d7652ecSHong Zhang   const PetscScalar *vals;
9145d7652ecSHong Zhang 
9155d7652ecSHong Zhang   PetscFunctionBegin;
9165d7652ecSHong Zhang   ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr);
9175d7652ecSHong Zhang   ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);CHKERRQ(ierr);
9185d7652ecSHong Zhang   ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr);
9195d7652ecSHong Zhang   ierr = MatSetUp(mat_elemental);CHKERRQ(ierr);
9205d7652ecSHong Zhang   for (row=rstart; row<rend; row++) {
9215d7652ecSHong Zhang     ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9225d7652ecSHong Zhang     for (j=0; j<ncols; j++) {
9235d7652ecSHong Zhang       /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
9245d7652ecSHong Zhang       ierr = MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);CHKERRQ(ierr);
9255d7652ecSHong Zhang     }
9265d7652ecSHong Zhang     ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr);
9275d7652ecSHong Zhang   }
9285d7652ecSHong Zhang   ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9295d7652ecSHong Zhang   ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
9305d7652ecSHong Zhang 
9315d7652ecSHong Zhang   if (reuse == MAT_REUSE_MATRIX) {
9325d7652ecSHong Zhang     ierr = MatHeaderReplace(A,mat_elemental);CHKERRQ(ierr);
9335d7652ecSHong Zhang   } else {
9345d7652ecSHong Zhang     *newmat = mat_elemental;
9355d7652ecSHong Zhang   }
9365d7652ecSHong Zhang   PetscFunctionReturn(0);
9375d7652ecSHong Zhang }
9385d7652ecSHong Zhang 
9395d7652ecSHong Zhang #undef __FUNCT__
940db31f6deSJed Brown #define __FUNCT__ "MatDestroy_Elemental"
941db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A)
942db31f6deSJed Brown {
943db31f6deSJed Brown   Mat_Elemental      *a = (Mat_Elemental*)A->data;
944db31f6deSJed Brown   PetscErrorCode     ierr;
9455e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
9465e9f5b67SHong Zhang   PetscBool          flg;
9475e9f5b67SHong Zhang   MPI_Comm           icomm;
948db31f6deSJed Brown 
949db31f6deSJed Brown   PetscFunctionBegin;
950c1ee1e62SHong Zhang   a->interface->Detach();
951aae2c449SHong Zhang   delete a->interface;
952aae2c449SHong Zhang   delete a->esubmat;
953db31f6deSJed Brown   delete a->emat;
9545e9f5b67SHong Zhang 
955ce94432eSBarry Smith   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
9560c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
9575e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
9585e9f5b67SHong Zhang   if (--commgrid->grid_refct == 0) {
9595e9f5b67SHong Zhang     delete commgrid->grid;
9605e9f5b67SHong Zhang     ierr = PetscFree(commgrid);CHKERRQ(ierr);
9615e9f5b67SHong Zhang   }
9625e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
963bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr);
964bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr);
965d98da988SHong Zhang   ierr = PetscObjectComposeFunction((PetscObject)A,"MatElementalHermitianGenDefiniteEig_C",NULL);CHKERRQ(ierr);
966db31f6deSJed Brown   ierr = PetscFree(A->data);CHKERRQ(ierr);
967db31f6deSJed Brown   PetscFunctionReturn(0);
968db31f6deSJed Brown }
969db31f6deSJed Brown 
970db31f6deSJed Brown #undef __FUNCT__
971db31f6deSJed Brown #define __FUNCT__ "MatSetUp_Elemental"
972db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A)
973db31f6deSJed Brown {
974db31f6deSJed Brown   Mat_Elemental  *a = (Mat_Elemental*)A->data;
975db31f6deSJed Brown   PetscErrorCode ierr;
976db31f6deSJed Brown   PetscMPIInt    rsize,csize;
977db31f6deSJed Brown 
978db31f6deSJed Brown   PetscFunctionBegin;
979db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
980db31f6deSJed Brown   ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
981db31f6deSJed Brown 
982efb79153SJack Poulson   a->emat->Resize(A->rmap->N,A->cmap->N);CHKERRQ(ierr);
983db31f6deSJed Brown   elem::Zero(*a->emat);
984db31f6deSJed Brown 
985db31f6deSJed Brown   ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr);
986db31f6deSJed Brown   ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr);
987ce94432eSBarry Smith   if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes");
988db31f6deSJed Brown   a->commsize = rsize;
989db31f6deSJed Brown   a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize;
990db31f6deSJed Brown   a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize;
991db31f6deSJed Brown   a->m[0]  = A->rmap->N / rsize + (a->mr[0] != rsize);
992db31f6deSJed Brown   a->m[1]  = A->cmap->N / csize + (a->mr[1] != csize);
993db31f6deSJed Brown   PetscFunctionReturn(0);
994db31f6deSJed Brown }
995db31f6deSJed Brown 
996aae2c449SHong Zhang #undef __FUNCT__
997aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyBegin_Elemental"
998aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type)
999aae2c449SHong Zhang {
1000aae2c449SHong Zhang   Mat_Elemental  *a = (Mat_Elemental*)A->data;
1001aae2c449SHong Zhang 
1002aae2c449SHong Zhang   PetscFunctionBegin;
1003aae2c449SHong Zhang   a->interface->Detach();
1004aae2c449SHong Zhang   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
1005aae2c449SHong Zhang   PetscFunctionReturn(0);
1006aae2c449SHong Zhang }
1007aae2c449SHong Zhang 
1008aae2c449SHong Zhang #undef __FUNCT__
1009aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyEnd_Elemental"
1010aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type)
1011aae2c449SHong Zhang {
1012aae2c449SHong Zhang   PetscFunctionBegin;
1013aae2c449SHong Zhang   /* Currently does nothing */
1014aae2c449SHong Zhang   PetscFunctionReturn(0);
1015aae2c449SHong Zhang }
1016aae2c449SHong Zhang 
10177b30e0f1SHong Zhang #undef __FUNCT__
10187b30e0f1SHong Zhang #define __FUNCT__ "MatLoad_Elemental"
10197b30e0f1SHong Zhang PetscErrorCode MatLoad_Elemental(Mat newMat, PetscViewer viewer)
10207b30e0f1SHong Zhang {
10217b30e0f1SHong Zhang   PetscErrorCode ierr;
10227b30e0f1SHong Zhang   Mat            Adense,Ae;
10237b30e0f1SHong Zhang   MPI_Comm       comm;
10247b30e0f1SHong Zhang 
10257b30e0f1SHong Zhang   PetscFunctionBegin;
10267b30e0f1SHong Zhang   ierr = PetscObjectGetComm((PetscObject)newMat,&comm);CHKERRQ(ierr);
10277b30e0f1SHong Zhang   ierr = MatCreate(comm,&Adense);CHKERRQ(ierr);
10287b30e0f1SHong Zhang   ierr = MatSetType(Adense,MATDENSE);CHKERRQ(ierr);
10297b30e0f1SHong Zhang   ierr = MatLoad(Adense,viewer);CHKERRQ(ierr);
10307b30e0f1SHong Zhang   ierr = MatConvert(Adense, MATELEMENTAL, MAT_INITIAL_MATRIX,&Ae);CHKERRQ(ierr);
10317b30e0f1SHong Zhang   ierr = MatDestroy(&Adense);CHKERRQ(ierr);
10327b30e0f1SHong Zhang   ierr = MatHeaderReplace(newMat,Ae);CHKERRQ(ierr);
10337b30e0f1SHong Zhang   PetscFunctionReturn(0);
10347b30e0f1SHong Zhang }
10357b30e0f1SHong Zhang 
10361d08bef3SHong Zhang #undef __FUNCT__
1037d98da988SHong Zhang #define __FUNCT__ "MatElementalHermitianGenDefiniteEig_Elemental"
1038*382fd914SHong Zhang PetscErrorCode MatElementalHermitianGenDefiniteEig_Elemental(elem::HermitianGenDefiniteEigType type,elem::UpperOrLower uplo1,Mat A,Mat B,Mat *evals,Mat *evec,PetscReal vl,PetscReal vu)
10391d08bef3SHong Zhang {
10405dd12673SHong Zhang   PetscErrorCode           ierr;
1041874e4880SHong Zhang   Mat_Elemental            *a=(Mat_Elemental*)A->data,*b=(Mat_Elemental*)B->data,*x;
1042d98da988SHong Zhang   PetscElemScalar          vle=(PetscElemScalar)vl,vue=(PetscElemScalar)vu;
1043d59fe0e1SHong Zhang   elem::HermitianGenDefiniteEigType eigtype = elem::AXBX;
10445dd12673SHong Zhang   const elem::UpperOrLower uplo = elem::UPPER;
10455dd12673SHong Zhang   const elem::SortType     sort = elem::UNSORTED; /* UNSORTED, DESCENDING, ASCENDING */
1046874e4880SHong Zhang   MPI_Comm                 comm;
1047874e4880SHong Zhang   Mat                      EVAL;
1048874e4880SHong Zhang 
10491d08bef3SHong Zhang   PetscFunctionBegin;
1050874e4880SHong Zhang   /* Compute eigenvalues and eigenvectors */
10515dd12673SHong Zhang   elem::DistMatrix<PetscElemScalar,elem::VR,elem::STAR> w( *a->grid ); /* holding eigenvalues */
1052f82baa17SHong Zhang   elem::DistMatrix<PetscElemScalar> X( *a->grid ); /* holding eigenvectors */
1053d59fe0e1SHong Zhang   elem::HermitianGenDefiniteEig(eigtype,uplo,*a->emat,*b->emat,w,X,vle,vue,sort);
10545dd12673SHong Zhang   /* elem::Print(w, "Eigenvalues"); */
10555dd12673SHong Zhang 
1056874e4880SHong Zhang   /* Wrap w and X into PETSc's MATMATELEMENTAL matrices */
1057d59fe0e1SHong Zhang   ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
1058d59fe0e1SHong Zhang   ierr = MatCreate(comm,evec);CHKERRQ(ierr);
1059d59fe0e1SHong Zhang   ierr = MatSetSizes(*evec,PETSC_DECIDE,PETSC_DECIDE,X.Height(),X.Width());CHKERRQ(ierr);
1060d59fe0e1SHong Zhang   ierr = MatSetType(*evec,MATELEMENTAL);CHKERRQ(ierr);
1061d59fe0e1SHong Zhang   ierr = MatSetFromOptions(*evec);CHKERRQ(ierr);
1062d59fe0e1SHong Zhang   ierr = MatSetUp(*evec);CHKERRQ(ierr);
1063d59fe0e1SHong Zhang   ierr = MatAssemblyBegin(*evec,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1064d59fe0e1SHong Zhang   ierr = MatAssemblyEnd(*evec,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1065d59fe0e1SHong Zhang 
1066874e4880SHong Zhang   x = (Mat_Elemental*)(*evec)->data;
1067d59fe0e1SHong Zhang   //delete x->emat; //-- memory leak???
1068d59fe0e1SHong Zhang   *x->emat = X;
1069d59fe0e1SHong Zhang 
1070874e4880SHong Zhang   ierr = MatCreate(comm,&EVAL);CHKERRQ(ierr);
1071874e4880SHong Zhang   ierr = MatSetSizes(EVAL,PETSC_DECIDE,PETSC_DECIDE,w.Height(),w.Width());CHKERRQ(ierr);
1072874e4880SHong Zhang   ierr = MatSetType(EVAL,MATELEMENTAL);CHKERRQ(ierr);
1073874e4880SHong Zhang   ierr = MatSetFromOptions(EVAL);CHKERRQ(ierr);
1074874e4880SHong Zhang   ierr = MatSetUp(EVAL);CHKERRQ(ierr);
1075874e4880SHong Zhang   ierr = MatAssemblyBegin(EVAL,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1076874e4880SHong Zhang   ierr = MatAssemblyEnd(EVAL,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
1077874e4880SHong Zhang   Mat_Elemental  *e = (Mat_Elemental*)EVAL->data;
1078874e4880SHong Zhang   *e->emat = w; //-- memory leak???
1079874e4880SHong Zhang   *evals   = EVAL;
1080874e4880SHong Zhang 
1081874e4880SHong Zhang #if defined(MV)
1082f82baa17SHong Zhang   /* Test correctness norm = || - A*X + B*X*w || */
1083d59fe0e1SHong Zhang   {
1084f82baa17SHong Zhang     PetscElemScalar alpha,beta;
1085f82baa17SHong Zhang     elem::DistMatrix<PetscElemScalar> Y(*a->grid); //tmp matrix
1086f82baa17SHong Zhang     alpha = 1.0; beta=0.0;
1087f82baa17SHong Zhang     elem::Gemm(elem::NORMAL,elem::NORMAL,alpha,*b->emat,X,beta,Y); //Y = B*X
1088f82baa17SHong Zhang     elem::DiagonalScale(elem::RIGHT,elem::NORMAL, w, Y); //Y = Y*w
1089f82baa17SHong Zhang     alpha = -1.0; beta=1.0;
1090f82baa17SHong Zhang     elem::Gemm(elem::NORMAL,elem::NORMAL,alpha,*a->emat,X,beta,Y); //Y = - A*X + B*X*w
1091f82baa17SHong Zhang 
1092f82baa17SHong Zhang     PetscElemScalar norm = elem::FrobeniusNorm(Y);
1093d59fe0e1SHong Zhang     if ((*a->grid).Rank()==0) printf("  norm (- A*X + B*X*w) = %g\n",norm);
1094d59fe0e1SHong Zhang   }
1095f82baa17SHong Zhang 
1096874e4880SHong Zhang   {
1097874e4880SHong Zhang     PetscMPIInt rank;
1098874e4880SHong Zhang     ierr = MPI_Comm_rank(comm,&rank);
1099874e4880SHong 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());
11005dd12673SHong Zhang   }
1101874e4880SHong Zhang #endif
11021d08bef3SHong Zhang   PetscFunctionReturn(0);
11031d08bef3SHong Zhang }
11041d08bef3SHong Zhang 
11051d08bef3SHong Zhang #undef __FUNCT__
1106d98da988SHong Zhang #define __FUNCT__ "MatElementalHermitianGenDefiniteEig"
11071d08bef3SHong Zhang /*@
1108d98da988SHong Zhang   MatElementalHermitianGenDefiniteEig -
11091d08bef3SHong Zhang 
11101d08bef3SHong Zhang    Logically Collective on Mat
11111d08bef3SHong Zhang 
11121d08bef3SHong Zhang    Input Parameters:
11131d08bef3SHong Zhang +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
11141d08bef3SHong Zhang .  icntl - index of MUMPS parameter array ICNTL()
11151d08bef3SHong Zhang -  ival - value of MUMPS ICNTL(icntl)
11161d08bef3SHong Zhang 
11171d08bef3SHong Zhang   Options Database:
11181d08bef3SHong Zhang .   -mat_mumps_icntl_<icntl> <ival>
11191d08bef3SHong Zhang 
11201d08bef3SHong Zhang    Level: beginner
11211d08bef3SHong Zhang 
11221d08bef3SHong Zhang    References: MUMPS Users' Guide
11231d08bef3SHong Zhang 
11241d08bef3SHong Zhang .seealso: MatGetFactor()
11251d08bef3SHong Zhang @*/
1126*382fd914SHong Zhang PetscErrorCode MatElementalHermitianGenDefiniteEig(elem::HermitianGenDefiniteEigType type,elem::UpperOrLower uplo,Mat A,Mat B,Mat *evals,Mat *evec,PetscReal vl,PetscReal vu)
11271d08bef3SHong Zhang {
11281d08bef3SHong Zhang   PetscErrorCode ierr;
11291d08bef3SHong Zhang 
11301d08bef3SHong Zhang   PetscFunctionBegin;
1131*382fd914SHong Zhang   ierr = PetscTryMethod(A,"MatElementalHermitianGenDefiniteEig_C",(elem::HermitianGenDefiniteEigType,elem::UpperOrLower,Mat,Mat,Mat*,Mat*,PetscReal,PetscReal),(type,uplo,A,B,evals,evec,vl,vu));CHKERRQ(ierr);
11321d08bef3SHong Zhang   PetscFunctionReturn(0);
11331d08bef3SHong Zhang }
11341d08bef3SHong Zhang 
113540d92e34SHong Zhang /* -------------------------------------------------------------------*/
113640d92e34SHong Zhang static struct _MatOps MatOps_Values = {
113740d92e34SHong Zhang        MatSetValues_Elemental,
113840d92e34SHong Zhang        0,
113940d92e34SHong Zhang        0,
114040d92e34SHong Zhang        MatMult_Elemental,
114140d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental,
11429426833fSXuan Zhou        MatMultTranspose_Elemental,
1143e883f9d5SXuan Zhou        MatMultTransposeAdd_Elemental,
114440d92e34SHong Zhang        MatSolve_Elemental,
1145df311e6cSXuan Zhou        MatSolveAdd_Elemental,
114640d92e34SHong Zhang        0, //MatSolveTranspose_Elemental,
114740d92e34SHong Zhang /*10*/ 0, //MatSolveTransposeAdd_Elemental,
114840d92e34SHong Zhang        MatLUFactor_Elemental,
114940d92e34SHong Zhang        MatCholeskyFactor_Elemental,
115040d92e34SHong Zhang        0,
115140d92e34SHong Zhang        MatTranspose_Elemental,
115240d92e34SHong Zhang /*15*/ MatGetInfo_Elemental,
115340d92e34SHong Zhang        0,
115461119200SXuan Zhou        MatGetDiagonal_Elemental,
1155ade3cc5eSXuan Zhou        MatDiagonalScale_Elemental,
115640d92e34SHong Zhang        MatNorm_Elemental,
115740d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental,
115840d92e34SHong Zhang        MatAssemblyEnd_Elemental,
115940d92e34SHong Zhang        0, //MatSetOption_Elemental,
116040d92e34SHong Zhang        MatZeroEntries_Elemental,
116140d92e34SHong Zhang /*24*/ 0,
116240d92e34SHong Zhang        MatLUFactorSymbolic_Elemental,
116340d92e34SHong Zhang        MatLUFactorNumeric_Elemental,
116440d92e34SHong Zhang        MatCholeskyFactorSymbolic_Elemental,
116540d92e34SHong Zhang        MatCholeskyFactorNumeric_Elemental,
116640d92e34SHong Zhang /*29*/ MatSetUp_Elemental,
116740d92e34SHong Zhang        0,
116840d92e34SHong Zhang        0,
116940d92e34SHong Zhang        0,
117040d92e34SHong Zhang        0,
1171df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental,
117240d92e34SHong Zhang        0,
117340d92e34SHong Zhang        0,
117440d92e34SHong Zhang        0,
117540d92e34SHong Zhang        0,
117640d92e34SHong Zhang /*39*/ MatAXPY_Elemental,
117740d92e34SHong Zhang        0,
117840d92e34SHong Zhang        0,
117940d92e34SHong Zhang        0,
118040d92e34SHong Zhang        MatCopy_Elemental,
118140d92e34SHong Zhang /*44*/ 0,
118240d92e34SHong Zhang        MatScale_Elemental,
118340d92e34SHong Zhang        0,
118440d92e34SHong Zhang        0,
118540d92e34SHong Zhang        0,
118640d92e34SHong Zhang /*49*/ 0,
118740d92e34SHong Zhang        0,
118840d92e34SHong Zhang        0,
118940d92e34SHong Zhang        0,
119040d92e34SHong Zhang        0,
119140d92e34SHong Zhang /*54*/ 0,
119240d92e34SHong Zhang        0,
119340d92e34SHong Zhang        0,
119440d92e34SHong Zhang        0,
119540d92e34SHong Zhang        0,
119640d92e34SHong Zhang /*59*/ 0,
119740d92e34SHong Zhang        MatDestroy_Elemental,
119840d92e34SHong Zhang        MatView_Elemental,
119940d92e34SHong Zhang        0,
120040d92e34SHong Zhang        0,
120140d92e34SHong Zhang /*64*/ 0,
120240d92e34SHong Zhang        0,
120340d92e34SHong Zhang        0,
120440d92e34SHong Zhang        0,
120540d92e34SHong Zhang        0,
120640d92e34SHong Zhang /*69*/ 0,
120740d92e34SHong Zhang        0,
12082ef0cf24SXuan Zhou        MatConvert_Elemental_Dense,
120940d92e34SHong Zhang        0,
121040d92e34SHong Zhang        0,
121140d92e34SHong Zhang /*74*/ 0,
121240d92e34SHong Zhang        0,
121340d92e34SHong Zhang        0,
121440d92e34SHong Zhang        0,
121540d92e34SHong Zhang        0,
121640d92e34SHong Zhang /*79*/ 0,
121740d92e34SHong Zhang        0,
121840d92e34SHong Zhang        0,
121940d92e34SHong Zhang        0,
12207b30e0f1SHong Zhang        MatLoad_Elemental,
122140d92e34SHong Zhang /*84*/ 0,
122240d92e34SHong Zhang        0,
122340d92e34SHong Zhang        0,
122440d92e34SHong Zhang        0,
122540d92e34SHong Zhang        0,
122640d92e34SHong Zhang /*89*/ MatMatMult_Elemental,
122740d92e34SHong Zhang        MatMatMultSymbolic_Elemental,
122840d92e34SHong Zhang        MatMatMultNumeric_Elemental,
122940d92e34SHong Zhang        0,
123040d92e34SHong Zhang        0,
123140d92e34SHong Zhang /*94*/ 0,
1232df311e6cSXuan Zhou        MatMatTransposeMult_Elemental,
1233df311e6cSXuan Zhou        MatMatTransposeMultSymbolic_Elemental,
1234df311e6cSXuan Zhou        MatMatTransposeMultNumeric_Elemental,
123540d92e34SHong Zhang        0,
123640d92e34SHong Zhang /*99*/ 0,
123740d92e34SHong Zhang        0,
123840d92e34SHong Zhang        0,
1239dfcb0403SXuan Zhou        MatConjugate_Elemental,
124040d92e34SHong Zhang        0,
124140d92e34SHong Zhang /*104*/0,
124240d92e34SHong Zhang        0,
124340d92e34SHong Zhang        0,
124440d92e34SHong Zhang        0,
124540d92e34SHong Zhang        0,
124640d92e34SHong Zhang /*109*/MatMatSolve_Elemental,
124740d92e34SHong Zhang        0,
124840d92e34SHong Zhang        0,
124940d92e34SHong Zhang        0,
125040d92e34SHong Zhang        0,
125140d92e34SHong Zhang /*114*/0,
125240d92e34SHong Zhang        0,
125340d92e34SHong Zhang        0,
125440d92e34SHong Zhang        0,
125540d92e34SHong Zhang        0,
125640d92e34SHong Zhang /*119*/0,
12574a29722dSXuan Zhou        MatHermitianTranspose_Elemental,
125840d92e34SHong Zhang        0,
125940d92e34SHong Zhang        0,
126040d92e34SHong Zhang        0,
126140d92e34SHong Zhang /*124*/0,
126240d92e34SHong Zhang        0,
126340d92e34SHong Zhang        0,
126440d92e34SHong Zhang        0,
126540d92e34SHong Zhang        0,
126640d92e34SHong Zhang /*129*/0,
126740d92e34SHong Zhang        0,
126840d92e34SHong Zhang        0,
126940d92e34SHong Zhang        0,
127040d92e34SHong Zhang        0,
127140d92e34SHong Zhang /*134*/0,
127240d92e34SHong Zhang        0,
127340d92e34SHong Zhang        0,
127440d92e34SHong Zhang        0,
127540d92e34SHong Zhang        0
127640d92e34SHong Zhang };
127740d92e34SHong Zhang 
1278ed36708cSHong Zhang /*MC
1279ed36708cSHong Zhang    MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package
1280ed36708cSHong Zhang 
1281ed36708cSHong Zhang    Options Database Keys:
12825cc86fc1SJed Brown + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions()
12835cc86fc1SJed Brown - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix
1284ed36708cSHong Zhang 
1285ed36708cSHong Zhang   Level: beginner
1286ed36708cSHong Zhang 
12875cb544a0SHong Zhang .seealso: MATDENSE
1288ed36708cSHong Zhang M*/
12894a29722dSXuan Zhou 
1290db31f6deSJed Brown #undef __FUNCT__
1291db31f6deSJed Brown #define __FUNCT__ "MatCreate_Elemental"
12928cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A)
1293db31f6deSJed Brown {
1294db31f6deSJed Brown   Mat_Elemental      *a;
1295db31f6deSJed Brown   PetscErrorCode     ierr;
12965682a260SJack Poulson   PetscBool          flg,flg1;
12975e9f5b67SHong Zhang   Mat_Elemental_Grid *commgrid;
12985e9f5b67SHong Zhang   MPI_Comm           icomm;
12995682a260SJack Poulson   PetscInt           optv1;
1300db31f6deSJed Brown 
1301db31f6deSJed Brown   PetscFunctionBegin;
1302607a6623SBarry Smith   ierr = PetscElementalInitializePackage();CHKERRQ(ierr);
130340d92e34SHong Zhang   ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr);
130440d92e34SHong Zhang   A->insertmode = NOT_SET_VALUES;
1305db31f6deSJed Brown 
1306b00a9115SJed Brown   ierr = PetscNewLog(A,&a);CHKERRQ(ierr);
1307db31f6deSJed Brown   A->data = (void*)a;
1308db31f6deSJed Brown 
1309db31f6deSJed Brown   /* Set up the elemental matrix */
1310ce94432eSBarry Smith   elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A));
13115e9f5b67SHong Zhang 
13125e9f5b67SHong Zhang   /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */
13135e9f5b67SHong Zhang   if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) {
1314180a43e4SHong Zhang     ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);
13155e9f5b67SHong Zhang   }
13160c18141cSBarry Smith   ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr);
13175e9f5b67SHong Zhang   ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr);
13185e9f5b67SHong Zhang   if (!flg) {
1319b00a9115SJed Brown     ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr);
13205cb544a0SHong Zhang 
1321ce94432eSBarry Smith     ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr);
13225cb544a0SHong Zhang     /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until elem::Grid() is called */
13230c18141cSBarry Smith     ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",elem::mpi::Size(cxxcomm),&optv1,&flg1);CHKERRQ(ierr);
13245682a260SJack Poulson     if (flg1) {
13250c18141cSBarry Smith       if (elem::mpi::Size(cxxcomm) % optv1 != 0) {
13260c18141cSBarry Smith         SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)elem::mpi::Size(cxxcomm));
1327ed667823SXuan Zhou       }
13285682a260SJack Poulson       commgrid->grid = new elem::Grid(cxxcomm,optv1); /* use user-provided grid height */
13292ef0cf24SXuan Zhou     } else {
13302adf0be3SHong Zhang       commgrid->grid = new elem::Grid(cxxcomm); /* use Elemental default grid sizes */
1331ed667823SXuan Zhou     }
13325e9f5b67SHong Zhang     commgrid->grid_refct = 1;
13335e9f5b67SHong Zhang     ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr);
13345cb544a0SHong Zhang     PetscOptionsEnd();
13355e9f5b67SHong Zhang   } else {
13365e9f5b67SHong Zhang     commgrid->grid_refct++;
13375e9f5b67SHong Zhang   }
13385e9f5b67SHong Zhang   ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr);
13395e9f5b67SHong Zhang   a->grid      = commgrid->grid;
1340df311e6cSXuan Zhou   a->emat      = new elem::DistMatrix<PetscElemScalar>(*a->grid);
1341df311e6cSXuan Zhou   a->esubmat   = new elem::Matrix<PetscElemScalar>(1,1);
1342df311e6cSXuan Zhou   a->interface = new elem::AxpyInterface<PetscElemScalar>;
13437c920d81SXuan Zhou   a->pivot     = new elem::DistMatrix<PetscInt,elem::VC,elem::STAR>;
1344db31f6deSJed Brown 
1345db31f6deSJed Brown   /* build cache for off array entries formed */
1346aae2c449SHong Zhang   a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat));
1347bafd5131SHong Zhang 
1348bdf89e91SBarry Smith   ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr);
1349d98da988SHong Zhang   ierr = PetscObjectComposeFunction((PetscObject)A,"MatElementalHermitianGenDefiniteEig_C",MatElementalHermitianGenDefiniteEig_Elemental);CHKERRQ(ierr);
1350db31f6deSJed Brown 
1351db31f6deSJed Brown   ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr);
1352db31f6deSJed Brown   PetscFunctionReturn(0);
1353db31f6deSJed Brown }
13544a29722dSXuan Zhou 
1355