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 Input Arguments: 170298fd71SBarry Smith . path - the dynamic library path or NULL 18db31f6deSJed Brown 19db31f6deSJed Brown Level: developer 20db31f6deSJed Brown 21db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalFinalizePackage() 22db31f6deSJed Brown @*/ 23db31f6deSJed Brown PetscErrorCode PetscElementalInitializePackage(const char *path) 24db31f6deSJed Brown { 25db31f6deSJed Brown PetscErrorCode ierr; 26db31f6deSJed Brown 27db31f6deSJed Brown PetscFunctionBegin; 28db31f6deSJed Brown if (elem::Initialized()) PetscFunctionReturn(0); 29db31f6deSJed 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 */ 30db31f6deSJed Brown int zero = 0; 31db31f6deSJed Brown char **nothing = 0; 32db31f6deSJed Brown elem::Initialize(zero,nothing); 33db31f6deSJed Brown } 34db31f6deSJed Brown ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr); 35db31f6deSJed Brown PetscFunctionReturn(0); 36db31f6deSJed Brown } 37db31f6deSJed Brown 38db31f6deSJed Brown #undef __FUNCT__ 39db31f6deSJed Brown #define __FUNCT__ "PetscElementalFinalizePackage" 40db31f6deSJed Brown /*@C 41db31f6deSJed Brown PetscElementalFinalizePackage - Finalize Elemental package 42db31f6deSJed Brown 43db31f6deSJed Brown Logically Collective 44db31f6deSJed Brown 45db31f6deSJed Brown Level: developer 46db31f6deSJed Brown 47db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage() 48db31f6deSJed Brown @*/ 49db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void) 50db31f6deSJed Brown { 51db31f6deSJed Brown 52db31f6deSJed Brown PetscFunctionBegin; 53db31f6deSJed Brown elem::Finalize(); 54db31f6deSJed Brown PetscFunctionReturn(0); 55db31f6deSJed Brown } 56db31f6deSJed Brown 57db31f6deSJed Brown #undef __FUNCT__ 58db31f6deSJed Brown #define __FUNCT__ "MatView_Elemental" 59db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer) 60db31f6deSJed Brown { 61db31f6deSJed Brown PetscErrorCode ierr; 62db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 63db31f6deSJed Brown PetscBool iascii; 64db31f6deSJed Brown 65db31f6deSJed Brown PetscFunctionBegin; 66db31f6deSJed Brown ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 67db31f6deSJed Brown if (iascii) { 68db31f6deSJed Brown PetscViewerFormat format; 69db31f6deSJed Brown ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 70db31f6deSJed Brown if (format == PETSC_VIEWER_ASCII_INFO) { 7179673f7bSHong Zhang /* call elemental viewing function */ 722d8adcc7SHong Zhang ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr); 73ed667823SXuan Zhou ierr = PetscViewerASCIIPrintf(viewer," allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr); 74ed667823SXuan Zhou ierr = PetscViewerASCIIPrintf(viewer," grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr); 754fe7bbcaSHong Zhang if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 7679673f7bSHong Zhang /* call elemental viewing function */ 77ce94432eSBarry Smith ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr); 784fe7bbcaSHong Zhang } 7979673f7bSHong Zhang 80db31f6deSJed Brown } else if (format == PETSC_VIEWER_DEFAULT) { 81db31f6deSJed Brown ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 82db31f6deSJed Brown ierr = PetscObjectPrintClassNamePrefixType((PetscObject)A,viewer,"Matrix Object");CHKERRQ(ierr); 83db31f6deSJed Brown a->emat->Print("Elemental matrix (cyclic ordering)"); 84db31f6deSJed Brown ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 85834d3fecSHong Zhang if (A->factortype == MAT_FACTOR_NONE){ 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); 91834d3fecSHong Zhang } 92ce94432eSBarry Smith } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format"); 93d2daa67eSHong Zhang } else { 945cb544a0SHong Zhang /* convert to dense format and call MatView() */ 9561119200SXuan Zhou Mat Adense; 96ce94432eSBarry Smith ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr); 9761119200SXuan Zhou ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr); 9861119200SXuan Zhou ierr = MatView(Adense,viewer);CHKERRQ(ierr); 9961119200SXuan Zhou ierr = MatDestroy(&Adense);CHKERRQ(ierr); 100d2daa67eSHong Zhang } 101db31f6deSJed Brown PetscFunctionReturn(0); 102db31f6deSJed Brown } 103db31f6deSJed Brown 104db31f6deSJed Brown #undef __FUNCT__ 105180a43e4SHong Zhang #define __FUNCT__ "MatGetInfo_Elemental" 10615767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info) 107180a43e4SHong Zhang { 10815767789SHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data; 10915767789SHong Zhang PetscMPIInt rank; 11015767789SHong Zhang 111180a43e4SHong Zhang PetscFunctionBegin; 112ce94432eSBarry Smith MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank); 11315767789SHong Zhang 11415767789SHong Zhang /* if (!rank) printf(" .........MatGetInfo_Elemental ...\n"); */ 1155cb544a0SHong Zhang info->block_size = 1.0; 11615767789SHong Zhang 11715767789SHong Zhang if (flag == MAT_LOCAL) { 11815767789SHong Zhang info->nz_allocated = (double)(*a->emat).AllocatedMemory(); /* locally allocated */ 11915767789SHong Zhang info->nz_used = info->nz_allocated; 12015767789SHong Zhang } else if (flag == MAT_GLOBAL_MAX) { 121ce94432eSBarry Smith //ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 12215767789SHong Zhang /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */ 12315767789SHong Zhang //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet"); 12415767789SHong Zhang } else if (flag == MAT_GLOBAL_SUM) { 12515767789SHong Zhang //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet"); 12615767789SHong Zhang info->nz_allocated = (double)(*a->emat).AllocatedMemory(); /* locally allocated */ 12715767789SHong Zhang info->nz_used = info->nz_allocated; /* assume Elemental does accurate allocation */ 128ce94432eSBarry Smith //ierr = MPI_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 12915767789SHong Zhang //PetscPrintf(PETSC_COMM_SELF," ... [%d] locally allocated %g\n",rank,info->nz_allocated); 13015767789SHong Zhang } 13115767789SHong Zhang 13215767789SHong Zhang info->nz_unneeded = 0.0; 13315767789SHong Zhang info->assemblies = (double)A->num_ass; 13415767789SHong Zhang info->mallocs = 0; 13515767789SHong Zhang info->memory = ((PetscObject)A)->mem; 13615767789SHong Zhang info->fill_ratio_given = 0; /* determined by Elemental */ 13715767789SHong Zhang info->fill_ratio_needed = 0; 13815767789SHong Zhang info->factor_mallocs = 0; 139180a43e4SHong Zhang PetscFunctionReturn(0); 140180a43e4SHong Zhang } 141180a43e4SHong Zhang 142180a43e4SHong Zhang #undef __FUNCT__ 143db31f6deSJed Brown #define __FUNCT__ "MatSetValues_Elemental" 144e6dea9dbSXuan Zhou static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode) 145db31f6deSJed Brown { 146db31f6deSJed Brown PetscErrorCode ierr; 147db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 148db31f6deSJed Brown PetscMPIInt rank; 149db31f6deSJed Brown PetscInt i,j,rrank,ridx,crank,cidx; 150db31f6deSJed Brown 151db31f6deSJed Brown PetscFunctionBegin; 152ce94432eSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 153db31f6deSJed Brown 154db31f6deSJed Brown const elem::Grid &grid = a->emat->Grid(); 155db31f6deSJed Brown for (i=0; i<nr; i++) { 156db31f6deSJed Brown PetscInt erow,ecol,elrow,elcol; 157db31f6deSJed Brown if (rows[i] < 0) continue; 158db31f6deSJed Brown P2RO(A,0,rows[i],&rrank,&ridx); 159db31f6deSJed Brown RO2E(A,0,rrank,ridx,&erow); 160ce94432eSBarry Smith if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation"); 161db31f6deSJed Brown for (j=0; j<nc; j++) { 162db31f6deSJed Brown if (cols[j] < 0) continue; 163db31f6deSJed Brown P2RO(A,1,cols[j],&crank,&cidx); 164db31f6deSJed Brown RO2E(A,1,crank,cidx,&ecol); 165ce94432eSBarry Smith if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation"); 166aae2c449SHong Zhang if (erow % grid.MCSize() != grid.MCRank() || ecol % grid.MRSize() != grid.MRRank()){ /* off-proc entry */ 167aae2c449SHong Zhang if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported"); 168aae2c449SHong 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); */ 169e6dea9dbSXuan Zhou a->esubmat->Set(0,0, (PetscElemScalar)vals[i*nc+j]); 170aae2c449SHong Zhang a->interface->Axpy(1.0,*(a->esubmat),erow,ecol); 171aae2c449SHong Zhang continue; 172ed36708cSHong Zhang } 173db31f6deSJed Brown elrow = erow / grid.MCSize(); 174db31f6deSJed Brown elcol = ecol / grid.MRSize(); 175db31f6deSJed Brown switch (imode) { 176e6dea9dbSXuan Zhou case INSERT_VALUES: a->emat->SetLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break; 177e6dea9dbSXuan Zhou case ADD_VALUES: a->emat->UpdateLocal(elrow,elcol,(PetscElemScalar)vals[i*nc+j]); break; 178ce94432eSBarry Smith default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode); 179db31f6deSJed Brown } 180db31f6deSJed Brown } 181db31f6deSJed Brown } 182db31f6deSJed Brown PetscFunctionReturn(0); 183db31f6deSJed Brown } 184db31f6deSJed Brown 185db31f6deSJed Brown #undef __FUNCT__ 186db31f6deSJed Brown #define __FUNCT__ "MatMult_Elemental" 187db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y) 188db31f6deSJed Brown { 189db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 190db31f6deSJed Brown PetscErrorCode ierr; 191e6dea9dbSXuan Zhou const PetscElemScalar *x; 192e6dea9dbSXuan Zhou PetscElemScalar *y; 193df311e6cSXuan Zhou PetscElemScalar one = 1,zero = 0; 194db31f6deSJed Brown 195db31f6deSJed Brown PetscFunctionBegin; 196e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 197e6dea9dbSXuan Zhou ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 198db31f6deSJed Brown { /* Scoping so that constructor is called before pointer is returned */ 199df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->cmap->N,1,0,x,A->cmap->n,*a->grid); 200df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ye(A->rmap->N,1,0,y,A->rmap->n,*a->grid); 201db31f6deSJed Brown elem::Gemv(elem::NORMAL,one,*a->emat,xe,zero,ye); 202db31f6deSJed Brown } 203e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 204e6dea9dbSXuan Zhou ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 205db31f6deSJed Brown PetscFunctionReturn(0); 206db31f6deSJed Brown } 207db31f6deSJed Brown 208db31f6deSJed Brown #undef __FUNCT__ 2099426833fSXuan Zhou #define __FUNCT__ "MatMultTranspose_Elemental" 2109426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y) 2119426833fSXuan Zhou { 2129426833fSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 2139426833fSXuan Zhou PetscErrorCode ierr; 214df311e6cSXuan Zhou const PetscElemScalar *x; 215df311e6cSXuan Zhou PetscElemScalar *y; 216e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 2179426833fSXuan Zhou 2189426833fSXuan Zhou PetscFunctionBegin; 219e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 220e6dea9dbSXuan Zhou ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 2219426833fSXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 222df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid); 223df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ye(A->cmap->N,1,0,y,A->cmap->n,*a->grid); 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 */ 246df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->cmap->N,1,0,x,A->cmap->n,*a->grid); 247df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ze(A->rmap->N,1,0,z,A->rmap->n,*a->grid); 248db31f6deSJed Brown elem::Gemv(elem::NORMAL,one,*a->emat,xe,one,ze); 249db31f6deSJed Brown } 250e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 251e6dea9dbSXuan Zhou ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 252db31f6deSJed Brown PetscFunctionReturn(0); 253db31f6deSJed Brown } 254db31f6deSJed Brown 255db31f6deSJed Brown #undef __FUNCT__ 256e883f9d5SXuan Zhou #define __FUNCT__ "MatMultTransposeAdd_Elemental" 257e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z) 258e883f9d5SXuan Zhou { 259e883f9d5SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 260e883f9d5SXuan Zhou PetscErrorCode ierr; 261df311e6cSXuan Zhou const PetscElemScalar *x; 262df311e6cSXuan Zhou PetscElemScalar *z; 263e6dea9dbSXuan Zhou PetscElemScalar one = 1; 264e883f9d5SXuan Zhou 265e883f9d5SXuan Zhou PetscFunctionBegin; 266e883f9d5SXuan Zhou if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);} 267e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 268e6dea9dbSXuan Zhou ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 269e883f9d5SXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 270df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid); 271df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> ze(A->cmap->N,1,0,z,A->cmap->n,*a->grid); 272e883f9d5SXuan Zhou elem::Gemv(elem::TRANSPOSE,one,*a->emat,xe,one,ze); 273e883f9d5SXuan Zhou } 274e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 275e6dea9dbSXuan Zhou ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 276e883f9d5SXuan Zhou PetscFunctionReturn(0); 277e883f9d5SXuan Zhou } 278e883f9d5SXuan Zhou 279e883f9d5SXuan Zhou #undef __FUNCT__ 2809a9e8502SHong Zhang #define __FUNCT__ "MatMatMultNumeric_Elemental" 2819a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C) 282c1d1b975SXuan Zhou { 283c1d1b975SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 284c1d1b975SXuan Zhou Mat_Elemental *b = (Mat_Elemental*)B->data; 2859a9e8502SHong Zhang Mat_Elemental *c = (Mat_Elemental*)C->data; 286e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 287c1d1b975SXuan Zhou 288c1d1b975SXuan Zhou PetscFunctionBegin; 289aae2c449SHong Zhang { /* Scoping so that constructor is called before pointer is returned */ 290c1d1b975SXuan Zhou elem::Gemm(elem::NORMAL,elem::NORMAL,one,*a->emat,*b->emat,zero,*c->emat); 291aae2c449SHong Zhang } 2929a9e8502SHong Zhang C->assembled = PETSC_TRUE; 2939a9e8502SHong Zhang PetscFunctionReturn(0); 2949a9e8502SHong Zhang } 2959a9e8502SHong Zhang 2969a9e8502SHong Zhang #undef __FUNCT__ 2979a9e8502SHong Zhang #define __FUNCT__ "MatMatMultSymbolic_Elemental" 2989a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C) 2999a9e8502SHong Zhang { 3009a9e8502SHong Zhang PetscErrorCode ierr; 3019a9e8502SHong Zhang Mat Ce; 302ce94432eSBarry Smith MPI_Comm comm; 3039a9e8502SHong Zhang 3049a9e8502SHong Zhang PetscFunctionBegin; 305ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 3069a9e8502SHong Zhang ierr = MatCreate(comm,&Ce);CHKERRQ(ierr); 3079a9e8502SHong Zhang ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 3089a9e8502SHong Zhang ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr); 3099a9e8502SHong Zhang ierr = MatSetUp(Ce);CHKERRQ(ierr); 3109a9e8502SHong Zhang *C = Ce; 3119a9e8502SHong Zhang PetscFunctionReturn(0); 3129a9e8502SHong Zhang } 3139a9e8502SHong Zhang 3149a9e8502SHong Zhang #undef __FUNCT__ 3159a9e8502SHong Zhang #define __FUNCT__ "MatMatMult_Elemental" 3169a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 3179a9e8502SHong Zhang { 3189a9e8502SHong Zhang PetscErrorCode ierr; 3199a9e8502SHong Zhang 3209a9e8502SHong Zhang PetscFunctionBegin; 3219a9e8502SHong Zhang if (scall == MAT_INITIAL_MATRIX){ 3229a9e8502SHong Zhang ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr); 3239a9e8502SHong Zhang } 3249a9e8502SHong Zhang ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr); 325c1d1b975SXuan Zhou PetscFunctionReturn(0); 326c1d1b975SXuan Zhou } 327c1d1b975SXuan Zhou 328c1d1b975SXuan Zhou #undef __FUNCT__ 329df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultNumeric_Elemental" 330df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C) 331df311e6cSXuan Zhou { 332df311e6cSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 333df311e6cSXuan Zhou Mat_Elemental *b = (Mat_Elemental*)B->data; 334df311e6cSXuan Zhou Mat_Elemental *c = (Mat_Elemental*)C->data; 335e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 336df311e6cSXuan Zhou 337df311e6cSXuan Zhou PetscFunctionBegin; 338df311e6cSXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 339df311e6cSXuan Zhou elem::Gemm(elem::NORMAL,elem::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat); 340df311e6cSXuan Zhou } 341df311e6cSXuan Zhou C->assembled = PETSC_TRUE; 342df311e6cSXuan Zhou PetscFunctionReturn(0); 343df311e6cSXuan Zhou } 344df311e6cSXuan Zhou 345df311e6cSXuan Zhou #undef __FUNCT__ 346df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMultSymbolic_Elemental" 347df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C) 348df311e6cSXuan Zhou { 349df311e6cSXuan Zhou PetscErrorCode ierr; 350df311e6cSXuan Zhou Mat Ce; 351ce94432eSBarry Smith MPI_Comm comm; 352df311e6cSXuan Zhou 353df311e6cSXuan Zhou PetscFunctionBegin; 354ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 355df311e6cSXuan Zhou ierr = MatCreate(comm,&Ce);CHKERRQ(ierr); 356df311e6cSXuan Zhou ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 357df311e6cSXuan Zhou ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr); 358df311e6cSXuan Zhou ierr = MatSetUp(Ce);CHKERRQ(ierr); 359df311e6cSXuan Zhou *C = Ce; 360df311e6cSXuan Zhou PetscFunctionReturn(0); 361df311e6cSXuan Zhou } 362df311e6cSXuan Zhou 363df311e6cSXuan Zhou #undef __FUNCT__ 364df311e6cSXuan Zhou #define __FUNCT__ "MatMatTransposeMult_Elemental" 365df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 366df311e6cSXuan Zhou { 367df311e6cSXuan Zhou PetscErrorCode ierr; 368df311e6cSXuan Zhou 369df311e6cSXuan Zhou PetscFunctionBegin; 370df311e6cSXuan Zhou if (scall == MAT_INITIAL_MATRIX){ 371df311e6cSXuan Zhou ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 372df311e6cSXuan Zhou ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr); 373df311e6cSXuan Zhou ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 374df311e6cSXuan Zhou } 375df311e6cSXuan Zhou ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 376df311e6cSXuan Zhou ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr); 377df311e6cSXuan Zhou ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 378df311e6cSXuan Zhou PetscFunctionReturn(0); 379df311e6cSXuan Zhou } 380df311e6cSXuan Zhou 381df311e6cSXuan Zhou #undef __FUNCT__ 38261119200SXuan Zhou #define __FUNCT__ "MatGetDiagonal_Elemental" 383a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D) 38461119200SXuan Zhou { 385a9d89745SXuan Zhou PetscInt i,nrows,ncols,nD,rrank,ridx,crank,cidx; 386a9d89745SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 38761119200SXuan Zhou PetscErrorCode ierr; 388a9d89745SXuan Zhou PetscElemScalar v; 389ce94432eSBarry Smith MPI_Comm comm; 39061119200SXuan Zhou 39161119200SXuan Zhou PetscFunctionBegin; 392ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 393a9d89745SXuan Zhou ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr); 394a9d89745SXuan Zhou nD = nrows>ncols ? ncols : nrows; 395a9d89745SXuan Zhou for (i=0; i<nD; i++) { 396a9d89745SXuan Zhou PetscInt erow,ecol; 397a9d89745SXuan Zhou P2RO(A,0,i,&rrank,&ridx); 398a9d89745SXuan Zhou RO2E(A,0,rrank,ridx,&erow); 399a9d89745SXuan Zhou if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation"); 400a9d89745SXuan Zhou P2RO(A,1,i,&crank,&cidx); 401a9d89745SXuan Zhou RO2E(A,1,crank,cidx,&ecol); 402a9d89745SXuan Zhou if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation"); 403a9d89745SXuan Zhou v = a->emat->Get(erow,ecol); 4047c8b904fSSatish Balay ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr); 405ade3cc5eSXuan Zhou } 406a9d89745SXuan Zhou ierr = VecAssemblyBegin(D);CHKERRQ(ierr); 407a9d89745SXuan Zhou ierr = VecAssemblyEnd(D);CHKERRQ(ierr); 40861119200SXuan Zhou PetscFunctionReturn(0); 40961119200SXuan Zhou } 41061119200SXuan Zhou 41161119200SXuan Zhou #undef __FUNCT__ 412ade3cc5eSXuan Zhou #define __FUNCT__ "MatDiagonalScale_Elemental" 413ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R) 414ade3cc5eSXuan Zhou { 415ade3cc5eSXuan Zhou Mat_Elemental *x = (Mat_Elemental*)X->data; 416ade3cc5eSXuan Zhou const PetscElemScalar *d; 417ade3cc5eSXuan Zhou PetscErrorCode ierr; 418ade3cc5eSXuan Zhou 419ade3cc5eSXuan Zhou PetscFunctionBegin; 4200298fd71SBarry Smith if (L == NULL) { 421ade3cc5eSXuan Zhou ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr); 422ade3cc5eSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de(X->cmap->N,1,0,d,X->cmap->n,*x->grid); 423ade3cc5eSXuan Zhou elem::DiagonalScale(elem::RIGHT,elem::NORMAL,de,*x->emat); 424ade3cc5eSXuan Zhou ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr); 425ade3cc5eSXuan Zhou } else { 426ade3cc5eSXuan Zhou ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr); 427ade3cc5eSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> de(X->rmap->N,1,0,d,X->rmap->n,*x->grid); 428ade3cc5eSXuan Zhou elem::DiagonalScale(elem::LEFT,elem::NORMAL,de,*x->emat); 429ade3cc5eSXuan Zhou ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr); 430ade3cc5eSXuan Zhou } 431ade3cc5eSXuan Zhou PetscFunctionReturn(0); 432ade3cc5eSXuan Zhou } 433ade3cc5eSXuan Zhou 434ade3cc5eSXuan Zhou #undef __FUNCT__ 4354ee44ac6SXuan Zhou #define __FUNCT__ "MatScale_Elemental" 436e6dea9dbSXuan Zhou static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a) 43765b78793SXuan Zhou { 43865b78793SXuan Zhou Mat_Elemental *x = (Mat_Elemental*)X->data; 43965b78793SXuan Zhou 44065b78793SXuan Zhou PetscFunctionBegin; 441e6dea9dbSXuan Zhou elem::Scal((PetscElemScalar)a,*x->emat); 44265b78793SXuan Zhou PetscFunctionReturn(0); 44365b78793SXuan Zhou } 44465b78793SXuan Zhou 44565b78793SXuan Zhou #undef __FUNCT__ 4464ee44ac6SXuan Zhou #define __FUNCT__ "MatAXPY_Elemental" 447e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str) 448e09a3074SHong Zhang { 449e09a3074SHong Zhang Mat_Elemental *x = (Mat_Elemental*)X->data; 450e09a3074SHong Zhang Mat_Elemental *y = (Mat_Elemental*)Y->data; 451e09a3074SHong Zhang 452e09a3074SHong Zhang PetscFunctionBegin; 453e6dea9dbSXuan Zhou elem::Axpy((PetscElemScalar)a,*x->emat,*y->emat); 454e09a3074SHong Zhang PetscFunctionReturn(0); 455e09a3074SHong Zhang } 456e09a3074SHong Zhang 457ae844d54SHong Zhang #undef __FUNCT__ 458d6223691SXuan Zhou #define __FUNCT__ "MatCopy_Elemental" 459d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str) 460d6223691SXuan Zhou { 461d6223691SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 462d6223691SXuan Zhou Mat_Elemental *b=(Mat_Elemental*)B->data; 463d6223691SXuan Zhou 464d6223691SXuan Zhou PetscFunctionBegin; 465d6223691SXuan Zhou elem::Copy(*a->emat,*b->emat); 466d6223691SXuan Zhou PetscFunctionReturn(0); 467d6223691SXuan Zhou } 468d6223691SXuan Zhou 469d6223691SXuan Zhou #undef __FUNCT__ 470df311e6cSXuan Zhou #define __FUNCT__ "MatDuplicate_Elemental" 471df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B) 472df311e6cSXuan Zhou { 473df311e6cSXuan Zhou Mat Be; 474ce94432eSBarry Smith MPI_Comm comm; 475df311e6cSXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 476df311e6cSXuan Zhou PetscErrorCode ierr; 477df311e6cSXuan Zhou 478df311e6cSXuan Zhou PetscFunctionBegin; 479ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 480df311e6cSXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 481df311e6cSXuan Zhou ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 482df311e6cSXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 483df311e6cSXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 484df311e6cSXuan Zhou *B = Be; 485df311e6cSXuan Zhou if (op == MAT_COPY_VALUES) { 486df311e6cSXuan Zhou Mat_Elemental *b=(Mat_Elemental*)Be->data; 487df311e6cSXuan Zhou elem::Copy(*a->emat,*b->emat); 488df311e6cSXuan Zhou } 489df311e6cSXuan Zhou Be->assembled = PETSC_TRUE; 490df311e6cSXuan Zhou PetscFunctionReturn(0); 491df311e6cSXuan Zhou } 492df311e6cSXuan Zhou 493df311e6cSXuan Zhou #undef __FUNCT__ 494d6223691SXuan Zhou #define __FUNCT__ "MatTranspose_Elemental" 495d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B) 496d6223691SXuan Zhou { 4975262d616SXuan Zhou Mat Be; 4985262d616SXuan Zhou PetscErrorCode ierr; 499ce94432eSBarry Smith MPI_Comm comm; 5004fe7bbcaSHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data, *b; 501d6223691SXuan Zhou 502d6223691SXuan Zhou PetscFunctionBegin; 503ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5045cb544a0SHong Zhang /* Only out-of-place supported */ 5055262d616SXuan Zhou if (reuse == MAT_INITIAL_MATRIX){ 5065262d616SXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 5075262d616SXuan Zhou ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 5085262d616SXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 5095262d616SXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 5105262d616SXuan Zhou *B = Be; 5115262d616SXuan Zhou } 5124fe7bbcaSHong Zhang b = (Mat_Elemental*)Be->data; 5135262d616SXuan Zhou elem::Transpose(*a->emat,*b->emat); 5145262d616SXuan Zhou Be->assembled = PETSC_TRUE; 515d6223691SXuan Zhou PetscFunctionReturn(0); 516d6223691SXuan Zhou } 517d6223691SXuan Zhou 518d6223691SXuan Zhou #undef __FUNCT__ 519dfcb0403SXuan Zhou #define __FUNCT__ "MatConjugate_Elemental" 520dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A) 521dfcb0403SXuan Zhou { 522dfcb0403SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 523dfcb0403SXuan Zhou 524dfcb0403SXuan Zhou PetscFunctionBegin; 525dfcb0403SXuan Zhou elem::Conjugate(*a->emat); 526dfcb0403SXuan Zhou PetscFunctionReturn(0); 527dfcb0403SXuan Zhou } 528dfcb0403SXuan Zhou 529dfcb0403SXuan Zhou #undef __FUNCT__ 5304a29722dSXuan Zhou #define __FUNCT__ "MatHermitianTranspose_Elemental" 5314a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B) 5324a29722dSXuan Zhou { 5334a29722dSXuan Zhou Mat Be; 5344a29722dSXuan Zhou PetscErrorCode ierr; 535ce94432eSBarry Smith MPI_Comm comm; 5364a29722dSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data, *b; 5374a29722dSXuan Zhou 5384a29722dSXuan Zhou PetscFunctionBegin; 539ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5405cb544a0SHong Zhang /* Only out-of-place supported */ 5414a29722dSXuan Zhou if (reuse == MAT_INITIAL_MATRIX){ 5424a29722dSXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 5434a29722dSXuan Zhou ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 5444a29722dSXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 5454a29722dSXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 5464a29722dSXuan Zhou *B = Be; 5474a29722dSXuan Zhou } 5484a29722dSXuan Zhou b = (Mat_Elemental*)Be->data; 5494a29722dSXuan Zhou elem::Adjoint(*a->emat,*b->emat); 5504a29722dSXuan Zhou Be->assembled = PETSC_TRUE; 5514a29722dSXuan Zhou PetscFunctionReturn(0); 5524a29722dSXuan Zhou } 5534a29722dSXuan Zhou 5544a29722dSXuan Zhou #undef __FUNCT__ 5551f881ff8SXuan Zhou #define __FUNCT__ "MatSolve_Elemental" 5561f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X) 5571f881ff8SXuan Zhou { 5581f881ff8SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 5591f881ff8SXuan Zhou PetscErrorCode ierr; 560df311e6cSXuan Zhou PetscElemScalar *x; 5611f881ff8SXuan Zhou 5621f881ff8SXuan Zhou PetscFunctionBegin; 56345cf121fSXuan Zhou ierr = VecCopy(B,X);CHKERRQ(ierr); 564e6dea9dbSXuan Zhou ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr); 565df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::VC,elem::STAR> xe(A->rmap->N,1,0,x,A->rmap->n,*a->grid); 566df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::MC,elem::MR> xer = xe; 567fc54b460SXuan Zhou switch (A->factortype) { 568fc54b460SXuan Zhou case MAT_FACTOR_LU: 5691f881ff8SXuan Zhou if ((*a->pivot).AllocatedMemory()) { 5704b28e8e8SXuan Zhou elem::SolveAfterLU(elem::NORMAL,*a->emat,*a->pivot,xer); 57145cf121fSXuan Zhou elem::Copy(xer,xe); 572fc54b460SXuan Zhou } else { 5734b28e8e8SXuan Zhou elem::SolveAfterLU(elem::NORMAL,*a->emat,xer); 57445cf121fSXuan Zhou elem::Copy(xer,xe); 5751f881ff8SXuan Zhou } 576fc54b460SXuan Zhou break; 577fc54b460SXuan Zhou case MAT_FACTOR_CHOLESKY: 578fc54b460SXuan Zhou elem::SolveAfterCholesky(elem::UPPER,elem::NORMAL,*a->emat,xer); 579fc54b460SXuan Zhou elem::Copy(xer,xe); 580fc54b460SXuan Zhou break; 581fc54b460SXuan Zhou default: 5824fe7bbcaSHong Zhang SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType"); 583fc54b460SXuan Zhou break; 5841f881ff8SXuan Zhou } 585e6dea9dbSXuan Zhou ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr); 5861f881ff8SXuan Zhou PetscFunctionReturn(0); 5871f881ff8SXuan Zhou } 5881f881ff8SXuan Zhou 5891f881ff8SXuan Zhou #undef __FUNCT__ 590df311e6cSXuan Zhou #define __FUNCT__ "MatSolveAdd_Elemental" 591df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X) 592df311e6cSXuan Zhou { 593df311e6cSXuan Zhou PetscErrorCode ierr; 594df311e6cSXuan Zhou 595df311e6cSXuan Zhou PetscFunctionBegin; 596df311e6cSXuan Zhou ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr); 5973d7f40dbSXuan Zhou ierr = VecAXPY(X,1,Y);CHKERRQ(ierr); 598df311e6cSXuan Zhou PetscFunctionReturn(0); 599df311e6cSXuan Zhou } 600df311e6cSXuan Zhou 601df311e6cSXuan Zhou #undef __FUNCT__ 602ae844d54SHong Zhang #define __FUNCT__ "MatMatSolve_Elemental" 603ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X) 604ae844d54SHong Zhang { 6051f0e42cfSHong Zhang Mat_Elemental *a=(Mat_Elemental*)A->data; 606d6223691SXuan Zhou Mat_Elemental *b=(Mat_Elemental*)B->data; 6071f0e42cfSHong Zhang Mat_Elemental *x=(Mat_Elemental*)X->data; 6081f0e42cfSHong Zhang 609ae844d54SHong Zhang PetscFunctionBegin; 610d6223691SXuan Zhou elem::Copy(*b->emat,*x->emat); 611fc54b460SXuan Zhou switch (A->factortype) { 612fc54b460SXuan Zhou case MAT_FACTOR_LU: 613d6223691SXuan Zhou if ((*a->pivot).AllocatedMemory()) { 6144b28e8e8SXuan Zhou elem::SolveAfterLU(elem::NORMAL,*a->emat,*a->pivot,*x->emat); 615fc54b460SXuan Zhou } else { 6164b28e8e8SXuan Zhou elem::SolveAfterLU(elem::NORMAL,*a->emat,*x->emat); 617d6223691SXuan Zhou } 618fc54b460SXuan Zhou break; 619fc54b460SXuan Zhou case MAT_FACTOR_CHOLESKY: 620fc54b460SXuan Zhou elem::SolveAfterCholesky(elem::UPPER,elem::NORMAL,*a->emat,*x->emat); 621fc54b460SXuan Zhou break; 622fc54b460SXuan Zhou default: 6234fe7bbcaSHong Zhang SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType"); 624fc54b460SXuan Zhou break; 625fc54b460SXuan Zhou } 626ae844d54SHong Zhang PetscFunctionReturn(0); 627ae844d54SHong Zhang } 628ae844d54SHong Zhang 629ae844d54SHong Zhang #undef __FUNCT__ 630ae844d54SHong Zhang #define __FUNCT__ "MatLUFactor_Elemental" 631ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info) 632ae844d54SHong Zhang { 6337c920d81SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 6347c920d81SXuan Zhou 635ae844d54SHong Zhang PetscFunctionBegin; 636d6223691SXuan Zhou if (info->dtcol){ 6377c920d81SXuan Zhou elem::LU(*a->emat,*a->pivot); 6381293973dSHong Zhang } else { 639d6223691SXuan Zhou elem::LU(*a->emat); 640d6223691SXuan Zhou } 6411293973dSHong Zhang A->factortype = MAT_FACTOR_LU; 642834d3fecSHong Zhang A->assembled = PETSC_TRUE; 643ae844d54SHong Zhang PetscFunctionReturn(0); 644ae844d54SHong Zhang } 645ae844d54SHong Zhang 646d7c3f9d8SHong Zhang #undef __FUNCT__ 647d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorNumeric_Elemental" 648d7c3f9d8SHong Zhang static PetscErrorCode MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info) 649d7c3f9d8SHong Zhang { 650d7c3f9d8SHong Zhang PetscErrorCode ierr; 651d7c3f9d8SHong Zhang 652d7c3f9d8SHong Zhang PetscFunctionBegin; 653d7c3f9d8SHong Zhang ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 654d7c3f9d8SHong Zhang ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr); 655d7c3f9d8SHong Zhang PetscFunctionReturn(0); 656d7c3f9d8SHong Zhang } 657d7c3f9d8SHong Zhang 658d7c3f9d8SHong Zhang #undef __FUNCT__ 659d7c3f9d8SHong Zhang #define __FUNCT__ "MatLUFactorSymbolic_Elemental" 660d7c3f9d8SHong Zhang static PetscErrorCode MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 661d7c3f9d8SHong Zhang { 662d7c3f9d8SHong Zhang PetscFunctionBegin; 663d7c3f9d8SHong Zhang /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */ 664d7c3f9d8SHong Zhang PetscFunctionReturn(0); 665d7c3f9d8SHong Zhang } 666d7c3f9d8SHong Zhang 66745cf121fSXuan Zhou #undef __FUNCT__ 66845cf121fSXuan Zhou #define __FUNCT__ "MatCholeskyFactor_Elemental" 66945cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info) 67045cf121fSXuan Zhou { 67145cf121fSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 672df311e6cSXuan Zhou elem::DistMatrix<PetscElemScalar,elem::MC,elem::STAR> d; 67345cf121fSXuan Zhou 67445cf121fSXuan Zhou PetscFunctionBegin; 675c9fc186eSXuan Zhou elem::Cholesky(elem::UPPER,*a->emat); 676fc54b460SXuan Zhou A->factortype = MAT_FACTOR_CHOLESKY; 677834d3fecSHong Zhang A->assembled = PETSC_TRUE; 67845cf121fSXuan Zhou PetscFunctionReturn(0); 67945cf121fSXuan Zhou } 68045cf121fSXuan Zhou 68179673f7bSHong Zhang #undef __FUNCT__ 68279673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorNumeric_Elemental" 68379673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info) 68479673f7bSHong Zhang { 685cb76c1d8SXuan Zhou PetscErrorCode ierr; 686cb76c1d8SXuan Zhou 687cb76c1d8SXuan Zhou PetscFunctionBegin; 688cb76c1d8SXuan Zhou ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 689cb76c1d8SXuan Zhou ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr); 690cb76c1d8SXuan Zhou PetscFunctionReturn(0); 69179673f7bSHong Zhang } 692cb76c1d8SXuan Zhou 69379673f7bSHong Zhang #undef __FUNCT__ 69479673f7bSHong Zhang #define __FUNCT__ "MatCholeskyFactorSymbolic_Elemental" 69579673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info) 69679673f7bSHong Zhang { 69779673f7bSHong Zhang PetscFunctionBegin; 69879673f7bSHong Zhang /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */ 69979673f7bSHong Zhang PetscFunctionReturn(0); 70079673f7bSHong Zhang } 70179673f7bSHong Zhang 7021293973dSHong Zhang #undef __FUNCT__ 70315767789SHong Zhang #define __FUNCT__ "MatFactorGetSolverPackage_elemental_elemental" 70415767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type) 70515767789SHong Zhang { 70615767789SHong Zhang PetscFunctionBegin; 70715767789SHong Zhang *type = MATSOLVERELEMENTAL; 70815767789SHong Zhang PetscFunctionReturn(0); 70915767789SHong Zhang } 71015767789SHong Zhang 71115767789SHong Zhang #undef __FUNCT__ 71215767789SHong Zhang #define __FUNCT__ "MatGetFactor_elemental_elemental" 71315767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F) 7141293973dSHong Zhang { 7151293973dSHong Zhang Mat B; 7161293973dSHong Zhang PetscErrorCode ierr; 7171293973dSHong Zhang 7181293973dSHong Zhang PetscFunctionBegin; 7191293973dSHong Zhang /* Create the factorization matrix */ 720ce94432eSBarry Smith ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 7211293973dSHong Zhang ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 7221293973dSHong Zhang ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr); 7231293973dSHong Zhang ierr = MatSetUp(B);CHKERRQ(ierr); 7241293973dSHong Zhang B->factortype = ftype; 72500de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C","MatFactorGetSolverPackage_elemental_elemental",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr); 7261293973dSHong Zhang *F = B; 7271293973dSHong Zhang PetscFunctionReturn(0); 7281293973dSHong Zhang } 7291293973dSHong Zhang 7301f881ff8SXuan Zhou #undef __FUNCT__ 7311f881ff8SXuan Zhou #define __FUNCT__ "MatNorm_Elemental" 7321f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm) 7331f881ff8SXuan Zhou { 7341f881ff8SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 7351f881ff8SXuan Zhou 7361f881ff8SXuan Zhou PetscFunctionBegin; 7371f881ff8SXuan Zhou switch (type){ 7381f881ff8SXuan Zhou case NORM_1: 7391f881ff8SXuan Zhou *nrm = elem::Norm(*a->emat,elem::ONE_NORM); 7401f881ff8SXuan Zhou break; 7411f881ff8SXuan Zhou case NORM_FROBENIUS: 7421f881ff8SXuan Zhou *nrm = elem::Norm(*a->emat,elem::FROBENIUS_NORM); 7431f881ff8SXuan Zhou break; 7441f881ff8SXuan Zhou case NORM_INFINITY: 7451f881ff8SXuan Zhou *nrm = elem::Norm(*a->emat,elem::INFINITY_NORM); 7461f881ff8SXuan Zhou break; 7471f881ff8SXuan Zhou default: 7481f881ff8SXuan Zhou printf("Error: unsupported norm type!\n"); 7491f881ff8SXuan Zhou } 7501f881ff8SXuan Zhou PetscFunctionReturn(0); 7511f881ff8SXuan Zhou } 7521f881ff8SXuan Zhou 7535262d616SXuan Zhou #undef __FUNCT__ 7545262d616SXuan Zhou #define __FUNCT__ "MatZeroEntries_Elemental" 7555262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A) 7565262d616SXuan Zhou { 7575262d616SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 7585262d616SXuan Zhou 7595262d616SXuan Zhou PetscFunctionBegin; 7605262d616SXuan Zhou elem::Zero(*a->emat); 7615262d616SXuan Zhou PetscFunctionReturn(0); 7625262d616SXuan Zhou } 7635262d616SXuan Zhou 764e09a3074SHong Zhang #undef __FUNCT__ 765db31f6deSJed Brown #define __FUNCT__ "MatGetOwnershipIS_Elemental" 766db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols) 767db31f6deSJed Brown { 768db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 769db31f6deSJed Brown PetscErrorCode ierr; 770db31f6deSJed Brown PetscInt i,m,shift,stride,*idx; 771db31f6deSJed Brown 772db31f6deSJed Brown PetscFunctionBegin; 773db31f6deSJed Brown if (rows) { 774db31f6deSJed Brown m = a->emat->LocalHeight(); 775db31f6deSJed Brown shift = a->emat->ColShift(); 776db31f6deSJed Brown stride = a->emat->ColStride(); 777db31f6deSJed Brown ierr = PetscMalloc(m*sizeof(PetscInt),&idx);CHKERRQ(ierr); 778db31f6deSJed Brown for (i=0; i<m; i++) { 779db31f6deSJed Brown PetscInt rank,offset; 780db31f6deSJed Brown E2RO(A,0,shift+i*stride,&rank,&offset); 781db31f6deSJed Brown RO2P(A,0,rank,offset,&idx[i]); 782db31f6deSJed Brown } 783db31f6deSJed Brown ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr); 784db31f6deSJed Brown } 785db31f6deSJed Brown if (cols) { 786db31f6deSJed Brown m = a->emat->LocalWidth(); 787db31f6deSJed Brown shift = a->emat->RowShift(); 788db31f6deSJed Brown stride = a->emat->RowStride(); 789db31f6deSJed Brown ierr = PetscMalloc(m*sizeof(PetscInt),&idx);CHKERRQ(ierr); 790db31f6deSJed Brown for (i=0; i<m; i++) { 791db31f6deSJed Brown PetscInt rank,offset; 792db31f6deSJed Brown E2RO(A,1,shift+i*stride,&rank,&offset); 793db31f6deSJed Brown RO2P(A,1,rank,offset,&idx[i]); 794db31f6deSJed Brown } 795db31f6deSJed Brown ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr); 796db31f6deSJed Brown } 797db31f6deSJed Brown PetscFunctionReturn(0); 798db31f6deSJed Brown } 799db31f6deSJed Brown 800db31f6deSJed Brown #undef __FUNCT__ 8012ef0cf24SXuan Zhou #define __FUNCT__ "MatConvert_Elemental_Dense" 80219fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B) 803af295397SXuan Zhou { 8042ef0cf24SXuan Zhou Mat Bmpi; 805af295397SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 806ce94432eSBarry Smith MPI_Comm comm; 8072ef0cf24SXuan Zhou PetscErrorCode ierr; 8082ef0cf24SXuan Zhou PetscInt rrank,ridx,crank,cidx,nrows,ncols,i,j; 809df311e6cSXuan Zhou PetscElemScalar v; 810573b0fb4SBarry Smith PetscBool s1,s2,s3; 811af295397SXuan Zhou 812af295397SXuan Zhou PetscFunctionBegin; 813ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 814573b0fb4SBarry Smith ierr = PetscStrcmp(newtype,MATDENSE,&s1);CHKERRQ(ierr); 815573b0fb4SBarry Smith ierr = PetscStrcmp(newtype,MATSEQDENSE,&s2);CHKERRQ(ierr); 816573b0fb4SBarry Smith ierr = PetscStrcmp(newtype,MATMPIDENSE,&s3);CHKERRQ(ierr); 817573b0fb4SBarry Smith if (!s1 && !s2 && !s3) SETERRQ(comm,PETSC_ERR_SUP,"Unsupported New MatType: must be MATDENSE, MATSEQDENSE or MATMPIDENSE"); 818af295397SXuan Zhou ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr); 819af295397SXuan Zhou ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 8202ef0cf24SXuan Zhou ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr); 821af295397SXuan Zhou ierr = MatSetUp(Bmpi);CHKERRQ(ierr); 8222ef0cf24SXuan Zhou ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr); 8232ef0cf24SXuan Zhou for (i=0; i<nrows; i++) { 8242ef0cf24SXuan Zhou PetscInt erow,ecol; 8252ef0cf24SXuan Zhou P2RO(A,0,i,&rrank,&ridx); 8262ef0cf24SXuan Zhou RO2E(A,0,rrank,ridx,&erow); 8272ef0cf24SXuan Zhou if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation"); 8282ef0cf24SXuan Zhou for (j=0; j<ncols; j++) { 8292ef0cf24SXuan Zhou P2RO(A,1,j,&crank,&cidx); 8302ef0cf24SXuan Zhou RO2E(A,1,crank,cidx,&ecol); 8312ef0cf24SXuan Zhou if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation"); 8322ef0cf24SXuan Zhou v = a->emat->Get(erow,ecol); 833e6dea9dbSXuan Zhou ierr = MatSetValues(Bmpi,1,&i,1,&j,(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr); 8342ef0cf24SXuan Zhou } 8352ef0cf24SXuan Zhou } 836af295397SXuan Zhou ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 837af295397SXuan Zhou ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 838c4ad791aSXuan Zhou if (reuse == MAT_REUSE_MATRIX) { 839c4ad791aSXuan Zhou ierr = MatHeaderReplace(A,Bmpi);CHKERRQ(ierr); 840c4ad791aSXuan Zhou } else { 841c4ad791aSXuan Zhou *B = Bmpi; 842c4ad791aSXuan Zhou } 843af295397SXuan Zhou PetscFunctionReturn(0); 844af295397SXuan Zhou } 845af295397SXuan Zhou 846af295397SXuan Zhou #undef __FUNCT__ 847db31f6deSJed Brown #define __FUNCT__ "MatDestroy_Elemental" 848db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A) 849db31f6deSJed Brown { 850db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 851db31f6deSJed Brown PetscErrorCode ierr; 8525e9f5b67SHong Zhang Mat_Elemental_Grid *commgrid; 8535e9f5b67SHong Zhang PetscBool flg; 8545e9f5b67SHong Zhang MPI_Comm icomm; 855db31f6deSJed Brown 856db31f6deSJed Brown PetscFunctionBegin; 857c1ee1e62SHong Zhang a->interface->Detach(); 858aae2c449SHong Zhang delete a->interface; 859aae2c449SHong Zhang delete a->esubmat; 860db31f6deSJed Brown delete a->emat; 8615e9f5b67SHong Zhang 862ce94432eSBarry Smith elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A)); 8630298fd71SBarry Smith ierr = PetscCommDuplicate(cxxcomm,&icomm,NULL);CHKERRQ(ierr); 8645e9f5b67SHong Zhang ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr); 8655e9f5b67SHong Zhang if (--commgrid->grid_refct == 0) { 8665e9f5b67SHong Zhang delete commgrid->grid; 8675e9f5b67SHong Zhang ierr = PetscFree(commgrid);CHKERRQ(ierr); 8685e9f5b67SHong Zhang } 8695e9f5b67SHong Zhang ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr); 87000de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C","",NULL);CHKERRQ(ierr); 87100de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetFactor_petsc_C","",NULL);CHKERRQ(ierr); 87200de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C","",NULL);CHKERRQ(ierr); 873db31f6deSJed Brown ierr = PetscFree(A->data);CHKERRQ(ierr); 874db31f6deSJed Brown PetscFunctionReturn(0); 875db31f6deSJed Brown } 876db31f6deSJed Brown 877db31f6deSJed Brown #undef __FUNCT__ 878db31f6deSJed Brown #define __FUNCT__ "MatSetUp_Elemental" 879db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A) 880db31f6deSJed Brown { 881db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 882db31f6deSJed Brown PetscErrorCode ierr; 883db31f6deSJed Brown PetscMPIInt rsize,csize; 884db31f6deSJed Brown 885db31f6deSJed Brown PetscFunctionBegin; 886db31f6deSJed Brown ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr); 887db31f6deSJed Brown ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr); 888db31f6deSJed Brown 889db31f6deSJed Brown a->emat->ResizeTo(A->rmap->N,A->cmap->N);CHKERRQ(ierr); 890db31f6deSJed Brown elem::Zero(*a->emat); 891db31f6deSJed Brown 892db31f6deSJed Brown ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr); 893db31f6deSJed Brown ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr); 894ce94432eSBarry Smith if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes"); 895db31f6deSJed Brown a->commsize = rsize; 896db31f6deSJed Brown a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize; 897db31f6deSJed Brown a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize; 898db31f6deSJed Brown a->m[0] = A->rmap->N / rsize + (a->mr[0] != rsize); 899db31f6deSJed Brown a->m[1] = A->cmap->N / csize + (a->mr[1] != csize); 900db31f6deSJed Brown PetscFunctionReturn(0); 901db31f6deSJed Brown } 902db31f6deSJed Brown 903aae2c449SHong Zhang #undef __FUNCT__ 904aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyBegin_Elemental" 905aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type) 906aae2c449SHong Zhang { 907aae2c449SHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data; 908aae2c449SHong Zhang 909aae2c449SHong Zhang PetscFunctionBegin; 910aae2c449SHong Zhang a->interface->Detach(); 911aae2c449SHong Zhang a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat)); 912aae2c449SHong Zhang PetscFunctionReturn(0); 913aae2c449SHong Zhang } 914aae2c449SHong Zhang 915aae2c449SHong Zhang #undef __FUNCT__ 916aae2c449SHong Zhang #define __FUNCT__ "MatAssemblyEnd_Elemental" 917aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type) 918aae2c449SHong Zhang { 919aae2c449SHong Zhang PetscFunctionBegin; 920aae2c449SHong Zhang /* Currently does nothing */ 921aae2c449SHong Zhang PetscFunctionReturn(0); 922aae2c449SHong Zhang } 923aae2c449SHong Zhang 92440d92e34SHong Zhang /* -------------------------------------------------------------------*/ 92540d92e34SHong Zhang static struct _MatOps MatOps_Values = { 92640d92e34SHong Zhang MatSetValues_Elemental, 92740d92e34SHong Zhang 0, 92840d92e34SHong Zhang 0, 92940d92e34SHong Zhang MatMult_Elemental, 93040d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental, 9319426833fSXuan Zhou MatMultTranspose_Elemental, 932e883f9d5SXuan Zhou MatMultTransposeAdd_Elemental, 93340d92e34SHong Zhang MatSolve_Elemental, 934df311e6cSXuan Zhou MatSolveAdd_Elemental, 93540d92e34SHong Zhang 0, //MatSolveTranspose_Elemental, 93640d92e34SHong Zhang /*10*/ 0, //MatSolveTransposeAdd_Elemental, 93740d92e34SHong Zhang MatLUFactor_Elemental, 93840d92e34SHong Zhang MatCholeskyFactor_Elemental, 93940d92e34SHong Zhang 0, 94040d92e34SHong Zhang MatTranspose_Elemental, 94140d92e34SHong Zhang /*15*/ MatGetInfo_Elemental, 94240d92e34SHong Zhang 0, 94361119200SXuan Zhou MatGetDiagonal_Elemental, 944ade3cc5eSXuan Zhou MatDiagonalScale_Elemental, 94540d92e34SHong Zhang MatNorm_Elemental, 94640d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental, 94740d92e34SHong Zhang MatAssemblyEnd_Elemental, 94840d92e34SHong Zhang 0, //MatSetOption_Elemental, 94940d92e34SHong Zhang MatZeroEntries_Elemental, 95040d92e34SHong Zhang /*24*/ 0, 95140d92e34SHong Zhang MatLUFactorSymbolic_Elemental, 95240d92e34SHong Zhang MatLUFactorNumeric_Elemental, 95340d92e34SHong Zhang MatCholeskyFactorSymbolic_Elemental, 95440d92e34SHong Zhang MatCholeskyFactorNumeric_Elemental, 95540d92e34SHong Zhang /*29*/ MatSetUp_Elemental, 95640d92e34SHong Zhang 0, 95740d92e34SHong Zhang 0, 95840d92e34SHong Zhang 0, 95940d92e34SHong Zhang 0, 960df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental, 96140d92e34SHong Zhang 0, 96240d92e34SHong Zhang 0, 96340d92e34SHong Zhang 0, 96440d92e34SHong Zhang 0, 96540d92e34SHong Zhang /*39*/ MatAXPY_Elemental, 96640d92e34SHong Zhang 0, 96740d92e34SHong Zhang 0, 96840d92e34SHong Zhang 0, 96940d92e34SHong Zhang MatCopy_Elemental, 97040d92e34SHong Zhang /*44*/ 0, 97140d92e34SHong Zhang MatScale_Elemental, 97240d92e34SHong Zhang 0, 97340d92e34SHong Zhang 0, 97440d92e34SHong Zhang 0, 97540d92e34SHong Zhang /*49*/ 0, 97640d92e34SHong Zhang 0, 97740d92e34SHong Zhang 0, 97840d92e34SHong Zhang 0, 97940d92e34SHong Zhang 0, 98040d92e34SHong Zhang /*54*/ 0, 98140d92e34SHong Zhang 0, 98240d92e34SHong Zhang 0, 98340d92e34SHong Zhang 0, 98440d92e34SHong Zhang 0, 98540d92e34SHong Zhang /*59*/ 0, 98640d92e34SHong Zhang MatDestroy_Elemental, 98740d92e34SHong Zhang MatView_Elemental, 98840d92e34SHong Zhang 0, 98940d92e34SHong Zhang 0, 99040d92e34SHong Zhang /*64*/ 0, 99140d92e34SHong Zhang 0, 99240d92e34SHong Zhang 0, 99340d92e34SHong Zhang 0, 99440d92e34SHong Zhang 0, 99540d92e34SHong Zhang /*69*/ 0, 99640d92e34SHong Zhang 0, 9972ef0cf24SXuan Zhou MatConvert_Elemental_Dense, 99840d92e34SHong Zhang 0, 99940d92e34SHong Zhang 0, 100040d92e34SHong Zhang /*74*/ 0, 100140d92e34SHong Zhang 0, 100240d92e34SHong Zhang 0, 100340d92e34SHong Zhang 0, 100440d92e34SHong Zhang 0, 100540d92e34SHong Zhang /*79*/ 0, 100640d92e34SHong Zhang 0, 100740d92e34SHong Zhang 0, 100840d92e34SHong Zhang 0, 100940d92e34SHong Zhang 0, 101040d92e34SHong Zhang /*84*/ 0, 101140d92e34SHong Zhang 0, 101240d92e34SHong Zhang 0, 101340d92e34SHong Zhang 0, 101440d92e34SHong Zhang 0, 101540d92e34SHong Zhang /*89*/ MatMatMult_Elemental, 101640d92e34SHong Zhang MatMatMultSymbolic_Elemental, 101740d92e34SHong Zhang MatMatMultNumeric_Elemental, 101840d92e34SHong Zhang 0, 101940d92e34SHong Zhang 0, 102040d92e34SHong Zhang /*94*/ 0, 1021df311e6cSXuan Zhou MatMatTransposeMult_Elemental, 1022df311e6cSXuan Zhou MatMatTransposeMultSymbolic_Elemental, 1023df311e6cSXuan Zhou MatMatTransposeMultNumeric_Elemental, 102440d92e34SHong Zhang 0, 102540d92e34SHong Zhang /*99*/ 0, 102640d92e34SHong Zhang 0, 102740d92e34SHong Zhang 0, 1028dfcb0403SXuan Zhou MatConjugate_Elemental, 102940d92e34SHong Zhang 0, 103040d92e34SHong Zhang /*104*/0, 103140d92e34SHong Zhang 0, 103240d92e34SHong Zhang 0, 103340d92e34SHong Zhang 0, 103440d92e34SHong Zhang 0, 103540d92e34SHong Zhang /*109*/MatMatSolve_Elemental, 103640d92e34SHong Zhang 0, 103740d92e34SHong Zhang 0, 103840d92e34SHong Zhang 0, 103940d92e34SHong Zhang 0, 104040d92e34SHong Zhang /*114*/0, 104140d92e34SHong Zhang 0, 104240d92e34SHong Zhang 0, 104340d92e34SHong Zhang 0, 104440d92e34SHong Zhang 0, 104540d92e34SHong Zhang /*119*/0, 10464a29722dSXuan Zhou MatHermitianTranspose_Elemental, 104740d92e34SHong Zhang 0, 104840d92e34SHong Zhang 0, 104940d92e34SHong Zhang 0, 105040d92e34SHong Zhang /*124*/0, 105140d92e34SHong Zhang 0, 105240d92e34SHong Zhang 0, 105340d92e34SHong Zhang 0, 105440d92e34SHong Zhang 0, 105540d92e34SHong Zhang /*129*/0, 105640d92e34SHong Zhang 0, 105740d92e34SHong Zhang 0, 105840d92e34SHong Zhang 0, 105940d92e34SHong Zhang 0, 106040d92e34SHong Zhang /*134*/0, 106140d92e34SHong Zhang 0, 106240d92e34SHong Zhang 0, 106340d92e34SHong Zhang 0, 106440d92e34SHong Zhang 0 106540d92e34SHong Zhang }; 106640d92e34SHong Zhang 1067ed36708cSHong Zhang /*MC 1068ed36708cSHong Zhang MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package 1069ed36708cSHong Zhang 1070ed36708cSHong Zhang Options Database Keys: 1071ed36708cSHong Zhang . -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions() 10725cb544a0SHong Zhang . -mat_elemental_grid_height - sets Grid Height 10735cb544a0SHong Zhang . -mat_elemental_grid_width - sets Grid Width 1074ed36708cSHong Zhang 1075ed36708cSHong Zhang Level: beginner 1076ed36708cSHong Zhang 10775cb544a0SHong Zhang .seealso: MATDENSE 1078ed36708cSHong Zhang M*/ 10794a29722dSXuan Zhou 1080db31f6deSJed Brown #undef __FUNCT__ 1081db31f6deSJed Brown #define __FUNCT__ "MatCreate_Elemental" 1082*8cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A) 1083db31f6deSJed Brown { 1084db31f6deSJed Brown Mat_Elemental *a; 1085db31f6deSJed Brown PetscErrorCode ierr; 1086ed667823SXuan Zhou PetscBool flg,flg1,flg2; 10875e9f5b67SHong Zhang Mat_Elemental_Grid *commgrid; 10885e9f5b67SHong Zhang MPI_Comm icomm; 1089ed667823SXuan Zhou PetscInt optv1,optv2; 1090db31f6deSJed Brown 1091db31f6deSJed Brown PetscFunctionBegin; 10920298fd71SBarry Smith ierr = PetscElementalInitializePackage(NULL);CHKERRQ(ierr); 109340d92e34SHong Zhang ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 109440d92e34SHong Zhang A->insertmode = NOT_SET_VALUES; 1095db31f6deSJed Brown 1096db31f6deSJed Brown ierr = PetscNewLog(A,Mat_Elemental,&a);CHKERRQ(ierr); 1097db31f6deSJed Brown A->data = (void*)a; 1098db31f6deSJed Brown 1099db31f6deSJed Brown /* Set up the elemental matrix */ 1100ce94432eSBarry Smith elem::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A)); 11015e9f5b67SHong Zhang 11025e9f5b67SHong Zhang /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */ 11035e9f5b67SHong Zhang if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) { 1104180a43e4SHong Zhang ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0); 11055e9f5b67SHong Zhang } 11060298fd71SBarry Smith ierr = PetscCommDuplicate(cxxcomm,&icomm,NULL);CHKERRQ(ierr); 11075e9f5b67SHong Zhang ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr); 11085e9f5b67SHong Zhang if (!flg) { 11095e9f5b67SHong Zhang ierr = PetscNewLog(A,Mat_Elemental_Grid,&commgrid);CHKERRQ(ierr); 11105cb544a0SHong Zhang 1111ce94432eSBarry Smith ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr); 11125cb544a0SHong Zhang /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until elem::Grid() is called */ 11132ef0cf24SXuan Zhou ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",elem::mpi::CommSize(cxxcomm),&optv1,&flg1);CHKERRQ(ierr); 11142ef0cf24SXuan Zhou ierr = PetscOptionsInt("-mat_elemental_grid_width","Grid Width","None",1,&optv2,&flg2);CHKERRQ(ierr); 1115ed667823SXuan Zhou if (flg1 || flg2) { 1116ed667823SXuan Zhou if (optv1*optv2 != elem::mpi::CommSize(cxxcomm)) { 1117ce94432eSBarry Smith SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height times Grid Width must equal CommSize"); 1118ed667823SXuan Zhou } 11192adf0be3SHong Zhang commgrid->grid = new elem::Grid(cxxcomm,optv1,optv2); /* use user-provided grid sizes */ 11202ef0cf24SXuan Zhou } else { 11212adf0be3SHong Zhang commgrid->grid = new elem::Grid(cxxcomm); /* use Elemental default grid sizes */ 1122ed667823SXuan Zhou } 11235e9f5b67SHong Zhang commgrid->grid_refct = 1; 11245e9f5b67SHong Zhang ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr); 11255cb544a0SHong Zhang PetscOptionsEnd(); 11265e9f5b67SHong Zhang } else { 11275e9f5b67SHong Zhang commgrid->grid_refct++; 11285e9f5b67SHong Zhang } 11295e9f5b67SHong Zhang ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr); 11305e9f5b67SHong Zhang a->grid = commgrid->grid; 1131df311e6cSXuan Zhou a->emat = new elem::DistMatrix<PetscElemScalar>(*a->grid); 1132df311e6cSXuan Zhou a->esubmat = new elem::Matrix<PetscElemScalar>(1,1); 1133df311e6cSXuan Zhou a->interface = new elem::AxpyInterface<PetscElemScalar>; 11347c920d81SXuan Zhou a->pivot = new elem::DistMatrix<PetscInt,elem::VC,elem::STAR>; 1135db31f6deSJed Brown 1136db31f6deSJed Brown /* build cache for off array entries formed */ 1137aae2c449SHong Zhang a->interface->Attach(elem::LOCAL_TO_GLOBAL,*(a->emat)); 1138bafd5131SHong Zhang 113900de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C","MatGetOwnershipIS_Elemental",MatGetOwnershipIS_Elemental);CHKERRQ(ierr); 114000de8ff0SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetFactor_elemental_C","MatGetFactor_elemental_elemental",MatGetFactor_elemental_elemental);CHKERRQ(ierr); 1141db31f6deSJed Brown 1142db31f6deSJed Brown ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr); 1143db31f6deSJed Brown PetscFunctionReturn(0); 1144db31f6deSJed Brown } 11454a29722dSXuan Zhou 1146