1db31f6deSJed Brown #include <../src/mat/impls/elemental/matelemimpl.h> /*I "petscmat.h" I*/ 2db31f6deSJed Brown 35e9f5b67SHong Zhang /* 45e9f5b67SHong Zhang The variable Petsc_Elemental_keyval is used to indicate an MPI attribute that 55e9f5b67SHong Zhang is attached to a communicator, in this case the attribute is a Mat_Elemental_Grid 65e9f5b67SHong Zhang */ 75e9f5b67SHong Zhang static PetscMPIInt Petsc_Elemental_keyval = MPI_KEYVAL_INVALID; 85e9f5b67SHong Zhang 9db31f6deSJed Brown /*@C 10db31f6deSJed Brown PetscElementalInitializePackage - Initialize Elemental package 11db31f6deSJed Brown 12db31f6deSJed Brown Logically Collective 13db31f6deSJed Brown 14db31f6deSJed Brown Level: developer 15db31f6deSJed Brown 16db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalFinalizePackage() 17db31f6deSJed Brown @*/ 18607a6623SBarry Smith PetscErrorCode PetscElementalInitializePackage(void) 19db31f6deSJed Brown { 20db31f6deSJed Brown PetscErrorCode ierr; 21db31f6deSJed Brown 22db31f6deSJed Brown PetscFunctionBegin; 23e8146892SSatish Balay if (El::Initialized()) PetscFunctionReturn(0); 24e05a9d59SSatish Balay El::Initialize(); /* called by the 1st call of MatCreate_Elemental */ 25db31f6deSJed Brown ierr = PetscRegisterFinalize(PetscElementalFinalizePackage);CHKERRQ(ierr); 26db31f6deSJed Brown PetscFunctionReturn(0); 27db31f6deSJed Brown } 28db31f6deSJed Brown 29db31f6deSJed Brown /*@C 30db31f6deSJed Brown PetscElementalFinalizePackage - Finalize Elemental package 31db31f6deSJed Brown 32db31f6deSJed Brown Logically Collective 33db31f6deSJed Brown 34db31f6deSJed Brown Level: developer 35db31f6deSJed Brown 36db31f6deSJed Brown .seealso: MATELEMENTAL, PetscElementalInitializePackage() 37db31f6deSJed Brown @*/ 38db31f6deSJed Brown PetscErrorCode PetscElementalFinalizePackage(void) 39db31f6deSJed Brown { 40db31f6deSJed Brown PetscFunctionBegin; 41e8146892SSatish Balay El::Finalize(); /* called by PetscFinalize() */ 42db31f6deSJed Brown PetscFunctionReturn(0); 43db31f6deSJed Brown } 44db31f6deSJed Brown 45db31f6deSJed Brown static PetscErrorCode MatView_Elemental(Mat A,PetscViewer viewer) 46db31f6deSJed Brown { 47db31f6deSJed Brown PetscErrorCode ierr; 48db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 49db31f6deSJed Brown PetscBool iascii; 50db31f6deSJed Brown 51db31f6deSJed Brown PetscFunctionBegin; 52db31f6deSJed Brown ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 53db31f6deSJed Brown if (iascii) { 54db31f6deSJed Brown PetscViewerFormat format; 55db31f6deSJed Brown ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 56db31f6deSJed Brown if (format == PETSC_VIEWER_ASCII_INFO) { 5779673f7bSHong Zhang /* call elemental viewing function */ 582d8adcc7SHong Zhang ierr = PetscViewerASCIIPrintf(viewer,"Elemental run parameters:\n");CHKERRQ(ierr); 59ed667823SXuan Zhou ierr = PetscViewerASCIIPrintf(viewer," allocated entries=%d\n",(*a->emat).AllocatedMemory());CHKERRQ(ierr); 60ed667823SXuan Zhou ierr = PetscViewerASCIIPrintf(viewer," grid height=%d, grid width=%d\n",(*a->emat).Grid().Height(),(*a->emat).Grid().Width());CHKERRQ(ierr); 614fe7bbcaSHong Zhang if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 6279673f7bSHong Zhang /* call elemental viewing function */ 63ce94432eSBarry Smith ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"test matview_elemental 2\n");CHKERRQ(ierr); 644fe7bbcaSHong Zhang } 6579673f7bSHong Zhang 66db31f6deSJed Brown } else if (format == PETSC_VIEWER_DEFAULT) { 67db31f6deSJed Brown ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 68e8146892SSatish Balay El::Print( *a->emat, "Elemental matrix (cyclic ordering)" ); 69db31f6deSJed Brown ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 70834d3fecSHong Zhang if (A->factortype == MAT_FACTOR_NONE){ 7161119200SXuan Zhou Mat Adense; 72ce94432eSBarry Smith ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr); 7361119200SXuan Zhou ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr); 7461119200SXuan Zhou ierr = MatView(Adense,viewer);CHKERRQ(ierr); 7561119200SXuan Zhou ierr = MatDestroy(&Adense);CHKERRQ(ierr); 76834d3fecSHong Zhang } 77ce94432eSBarry Smith } else SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Format"); 78d2daa67eSHong Zhang } else { 795cb544a0SHong Zhang /* convert to dense format and call MatView() */ 8061119200SXuan Zhou Mat Adense; 81ce94432eSBarry Smith ierr = PetscPrintf(PetscObjectComm((PetscObject)viewer),"Elemental matrix (explicit ordering)\n");CHKERRQ(ierr); 8261119200SXuan Zhou ierr = MatConvert(A,MATDENSE,MAT_INITIAL_MATRIX,&Adense);CHKERRQ(ierr); 8361119200SXuan Zhou ierr = MatView(Adense,viewer);CHKERRQ(ierr); 8461119200SXuan Zhou ierr = MatDestroy(&Adense);CHKERRQ(ierr); 85d2daa67eSHong Zhang } 86db31f6deSJed Brown PetscFunctionReturn(0); 87db31f6deSJed Brown } 88db31f6deSJed Brown 8915767789SHong Zhang static PetscErrorCode MatGetInfo_Elemental(Mat A,MatInfoType flag,MatInfo *info) 90180a43e4SHong Zhang { 9115767789SHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data; 9215767789SHong Zhang 93180a43e4SHong Zhang PetscFunctionBegin; 945cb544a0SHong Zhang info->block_size = 1.0; 9515767789SHong Zhang 9615767789SHong Zhang if (flag == MAT_LOCAL) { 9715767789SHong Zhang info->nz_allocated = (double)(*a->emat).AllocatedMemory(); /* locally allocated */ 9815767789SHong Zhang info->nz_used = info->nz_allocated; 9915767789SHong Zhang } else if (flag == MAT_GLOBAL_MAX) { 100b2566f29SBarry Smith //ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 10115767789SHong Zhang /* see MatGetInfo_MPIAIJ() for getting global info->nz_allocated! */ 10215767789SHong Zhang //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_MAX not written yet"); 10315767789SHong Zhang } else if (flag == MAT_GLOBAL_SUM) { 10415767789SHong Zhang //SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP," MAT_GLOBAL_SUM not written yet"); 10515767789SHong Zhang info->nz_allocated = (double)(*a->emat).AllocatedMemory(); /* locally allocated */ 10615767789SHong Zhang info->nz_used = info->nz_allocated; /* assume Elemental does accurate allocation */ 107b2566f29SBarry Smith //ierr = MPIU_Allreduce(isend,irecv,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 10815767789SHong Zhang //PetscPrintf(PETSC_COMM_SELF," ... [%d] locally allocated %g\n",rank,info->nz_allocated); 10915767789SHong Zhang } 11015767789SHong Zhang 11115767789SHong Zhang info->nz_unneeded = 0.0; 11215767789SHong Zhang info->assemblies = (double)A->num_ass; 11315767789SHong Zhang info->mallocs = 0; 11415767789SHong Zhang info->memory = ((PetscObject)A)->mem; 11515767789SHong Zhang info->fill_ratio_given = 0; /* determined by Elemental */ 11615767789SHong Zhang info->fill_ratio_needed = 0; 11715767789SHong Zhang info->factor_mallocs = 0; 118180a43e4SHong Zhang PetscFunctionReturn(0); 119180a43e4SHong Zhang } 120180a43e4SHong Zhang 12132d7a744SHong Zhang PetscErrorCode MatSetOption_Elemental(Mat A,MatOption op,PetscBool flg) 12232d7a744SHong Zhang { 12332d7a744SHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data; 12432d7a744SHong Zhang 12532d7a744SHong Zhang PetscFunctionBegin; 12632d7a744SHong Zhang switch (op) { 12732d7a744SHong Zhang case MAT_NEW_NONZERO_LOCATIONS: 12832d7a744SHong Zhang case MAT_NEW_NONZERO_LOCATION_ERR: 12932d7a744SHong Zhang case MAT_NEW_NONZERO_ALLOCATION_ERR: 130891a0571SHong Zhang case MAT_SYMMETRIC: 131891a0571SHong Zhang break; 13232d7a744SHong Zhang case MAT_ROW_ORIENTED: 13332d7a744SHong Zhang a->roworiented = flg; 13432d7a744SHong Zhang break; 13532d7a744SHong Zhang default: 13632d7a744SHong Zhang SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %s",MatOptions[op]); 13732d7a744SHong Zhang } 13832d7a744SHong Zhang PetscFunctionReturn(0); 13932d7a744SHong Zhang } 14032d7a744SHong Zhang 141e6dea9dbSXuan Zhou static PetscErrorCode MatSetValues_Elemental(Mat A,PetscInt nr,const PetscInt *rows,PetscInt nc,const PetscInt *cols,const PetscScalar *vals,InsertMode imode) 142db31f6deSJed Brown { 143db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 144fbbcb730SJack Poulson PetscInt i,j,rrank,ridx,crank,cidx,erow,ecol,numQueues=0; 145db31f6deSJed Brown 146db31f6deSJed Brown PetscFunctionBegin; 147fbbcb730SJack Poulson // TODO: Initialize matrix to all zeros? 148fbbcb730SJack Poulson 149fbbcb730SJack Poulson // Count the number of queues from this process 15032d7a744SHong Zhang if (a->roworiented) { 151db31f6deSJed Brown for (i=0; i<nr; i++) { 152db31f6deSJed Brown if (rows[i] < 0) continue; 153db31f6deSJed Brown P2RO(A,0,rows[i],&rrank,&ridx); 154db31f6deSJed Brown RO2E(A,0,rrank,ridx,&erow); 155ce94432eSBarry Smith if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation"); 156db31f6deSJed Brown for (j=0; j<nc; j++) { 157db31f6deSJed Brown if (cols[j] < 0) continue; 158db31f6deSJed Brown P2RO(A,1,cols[j],&crank,&cidx); 159db31f6deSJed Brown RO2E(A,1,crank,cidx,&ecol); 160ce94432eSBarry Smith if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation"); 161fbbcb730SJack Poulson if (!a->emat->IsLocal(erow,ecol) ){ /* off-proc entry */ 162fbbcb730SJack Poulson /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */ 163aae2c449SHong Zhang if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported"); 164fbbcb730SJack Poulson ++numQueues; 165aae2c449SHong Zhang continue; 166ed36708cSHong Zhang } 167fbbcb730SJack Poulson /* printf("Locally updating (%d,%d)\n",erow,ecol); */ 168db31f6deSJed Brown switch (imode) { 169fbbcb730SJack Poulson case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break; 170fbbcb730SJack Poulson case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i*nc+j]); break; 171ce94432eSBarry Smith default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode); 172db31f6deSJed Brown } 173db31f6deSJed Brown } 174db31f6deSJed Brown } 175fbbcb730SJack Poulson 176fbbcb730SJack Poulson /* printf("numQueues=%d\n",numQueues); */ 177fbbcb730SJack Poulson a->emat->Reserve( numQueues ); 178fbbcb730SJack Poulson for (i=0; i<nr; i++) { 179fbbcb730SJack Poulson if (rows[i] < 0) continue; 180fbbcb730SJack Poulson P2RO(A,0,rows[i],&rrank,&ridx); 181fbbcb730SJack Poulson RO2E(A,0,rrank,ridx,&erow); 182fbbcb730SJack Poulson for (j=0; j<nc; j++) { 183fbbcb730SJack Poulson if (cols[j] < 0) continue; 184fbbcb730SJack Poulson P2RO(A,1,cols[j],&crank,&cidx); 185fbbcb730SJack Poulson RO2E(A,1,crank,cidx,&ecol); 186fbbcb730SJack Poulson if ( !a->emat->IsLocal(erow,ecol) ) { /*off-proc entry*/ 187fbbcb730SJack Poulson /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */ 188fbbcb730SJack Poulson a->emat->QueueUpdate( erow, ecol, vals[i*nc+j] ); 189fbbcb730SJack Poulson } 190fbbcb730SJack Poulson } 191fbbcb730SJack Poulson } 19232d7a744SHong Zhang } else { /* columnoriented */ 19332d7a744SHong Zhang for (j=0; j<nc; j++) { 19432d7a744SHong Zhang if (cols[j] < 0) continue; 19532d7a744SHong Zhang P2RO(A,1,cols[j],&crank,&cidx); 19632d7a744SHong Zhang RO2E(A,1,crank,cidx,&ecol); 19732d7a744SHong Zhang if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect col translation"); 19832d7a744SHong Zhang for (i=0; i<nr; i++) { 19932d7a744SHong Zhang if (rows[i] < 0) continue; 20032d7a744SHong Zhang P2RO(A,0,rows[i],&rrank,&ridx); 20132d7a744SHong Zhang RO2E(A,0,rrank,ridx,&erow); 20232d7a744SHong Zhang if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Incorrect row translation"); 20332d7a744SHong Zhang if (!a->emat->IsLocal(erow,ecol) ){ /* off-proc entry */ 20432d7a744SHong Zhang /* printf("Will later remotely update (%d,%d)\n",erow,ecol); */ 20532d7a744SHong Zhang if (imode != ADD_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only ADD_VALUES to off-processor entry is supported"); 20632d7a744SHong Zhang ++numQueues; 20732d7a744SHong Zhang continue; 20832d7a744SHong Zhang } 20932d7a744SHong Zhang /* printf("Locally updating (%d,%d)\n",erow,ecol); */ 21032d7a744SHong Zhang switch (imode) { 21132d7a744SHong Zhang case INSERT_VALUES: a->emat->Set(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break; 21232d7a744SHong Zhang case ADD_VALUES: a->emat->Update(erow,ecol,(PetscElemScalar)vals[i+j*nr]); break; 21332d7a744SHong Zhang default: SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for InsertMode %d",(int)imode); 21432d7a744SHong Zhang } 21532d7a744SHong Zhang } 21632d7a744SHong Zhang } 21732d7a744SHong Zhang 21832d7a744SHong Zhang /* printf("numQueues=%d\n",numQueues); */ 21932d7a744SHong Zhang a->emat->Reserve( numQueues ); 22032d7a744SHong Zhang for (j=0; j<nc; j++) { 22132d7a744SHong Zhang if (cols[j] < 0) continue; 22232d7a744SHong Zhang P2RO(A,1,cols[j],&crank,&cidx); 22332d7a744SHong Zhang RO2E(A,1,crank,cidx,&ecol); 22432d7a744SHong Zhang 22532d7a744SHong Zhang for (i=0; i<nr; i++) { 22632d7a744SHong Zhang if (rows[i] < 0) continue; 22732d7a744SHong Zhang P2RO(A,0,rows[i],&rrank,&ridx); 22832d7a744SHong Zhang RO2E(A,0,rrank,ridx,&erow); 22932d7a744SHong Zhang if ( !a->emat->IsLocal(erow,ecol) ) { /*off-proc entry*/ 23032d7a744SHong Zhang /* printf("Queueing remotely update of (%d,%d)\n",erow,ecol); */ 23132d7a744SHong Zhang a->emat->QueueUpdate( erow, ecol, vals[i+j*nr] ); 23232d7a744SHong Zhang } 23332d7a744SHong Zhang } 23432d7a744SHong Zhang } 23532d7a744SHong Zhang } 236db31f6deSJed Brown PetscFunctionReturn(0); 237db31f6deSJed Brown } 238db31f6deSJed Brown 239db31f6deSJed Brown static PetscErrorCode MatMult_Elemental(Mat A,Vec X,Vec Y) 240db31f6deSJed Brown { 241db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 242db31f6deSJed Brown PetscErrorCode ierr; 243e6dea9dbSXuan Zhou const PetscElemScalar *x; 244e6dea9dbSXuan Zhou PetscElemScalar *y; 245df311e6cSXuan Zhou PetscElemScalar one = 1,zero = 0; 246db31f6deSJed Brown 247db31f6deSJed Brown PetscFunctionBegin; 248e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 249e6dea9dbSXuan Zhou ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 250db31f6deSJed Brown { /* Scoping so that constructor is called before pointer is returned */ 251e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye; 2520c18141cSBarry Smith xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n); 2530c18141cSBarry Smith ye.Attach(A->rmap->N,1,*a->grid,0,0,y,A->rmap->n); 254e8146892SSatish Balay El::Gemv(El::NORMAL,one,*a->emat,xe,zero,ye); 255db31f6deSJed Brown } 256e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 257e6dea9dbSXuan Zhou ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 258db31f6deSJed Brown PetscFunctionReturn(0); 259db31f6deSJed Brown } 260db31f6deSJed Brown 2619426833fSXuan Zhou static PetscErrorCode MatMultTranspose_Elemental(Mat A,Vec X,Vec Y) 2629426833fSXuan Zhou { 2639426833fSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 2649426833fSXuan Zhou PetscErrorCode ierr; 265df311e6cSXuan Zhou const PetscElemScalar *x; 266df311e6cSXuan Zhou PetscElemScalar *y; 267e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 2689426833fSXuan Zhou 2699426833fSXuan Zhou PetscFunctionBegin; 270e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 271e6dea9dbSXuan Zhou ierr = VecGetArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 2729426833fSXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 273e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ye; 2740c18141cSBarry Smith xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n); 2750c18141cSBarry Smith ye.Attach(A->cmap->N,1,*a->grid,0,0,y,A->cmap->n); 276e8146892SSatish Balay El::Gemv(El::TRANSPOSE,one,*a->emat,xe,zero,ye); 2779426833fSXuan Zhou } 278e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 279e6dea9dbSXuan Zhou ierr = VecRestoreArray(Y,(PetscScalar **)&y);CHKERRQ(ierr); 2809426833fSXuan Zhou PetscFunctionReturn(0); 2819426833fSXuan Zhou } 2829426833fSXuan Zhou 283db31f6deSJed Brown static PetscErrorCode MatMultAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z) 284db31f6deSJed Brown { 285db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 286db31f6deSJed Brown PetscErrorCode ierr; 287df311e6cSXuan Zhou const PetscElemScalar *x; 288df311e6cSXuan Zhou PetscElemScalar *z; 289e6dea9dbSXuan Zhou PetscElemScalar one = 1; 290db31f6deSJed Brown 291db31f6deSJed Brown PetscFunctionBegin; 292db31f6deSJed Brown if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);} 293e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 294e6dea9dbSXuan Zhou ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 295db31f6deSJed Brown { /* Scoping so that constructor is called before pointer is returned */ 296e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze; 2970c18141cSBarry Smith xe.LockedAttach(A->cmap->N,1,*a->grid,0,0,x,A->cmap->n); 2980c18141cSBarry Smith ze.Attach(A->rmap->N,1,*a->grid,0,0,z,A->rmap->n); 299e8146892SSatish Balay El::Gemv(El::NORMAL,one,*a->emat,xe,one,ze); 300db31f6deSJed Brown } 301e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 302e6dea9dbSXuan Zhou ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 303db31f6deSJed Brown PetscFunctionReturn(0); 304db31f6deSJed Brown } 305db31f6deSJed Brown 306e883f9d5SXuan Zhou static PetscErrorCode MatMultTransposeAdd_Elemental(Mat A,Vec X,Vec Y,Vec Z) 307e883f9d5SXuan Zhou { 308e883f9d5SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 309e883f9d5SXuan Zhou PetscErrorCode ierr; 310df311e6cSXuan Zhou const PetscElemScalar *x; 311df311e6cSXuan Zhou PetscElemScalar *z; 312e6dea9dbSXuan Zhou PetscElemScalar one = 1; 313e883f9d5SXuan Zhou 314e883f9d5SXuan Zhou PetscFunctionBegin; 315e883f9d5SXuan Zhou if (Y != Z) {ierr = VecCopy(Y,Z);CHKERRQ(ierr);} 316e6dea9dbSXuan Zhou ierr = VecGetArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 317e6dea9dbSXuan Zhou ierr = VecGetArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 318e883f9d5SXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 319e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe, ze; 3200c18141cSBarry Smith xe.LockedAttach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n); 3210c18141cSBarry Smith ze.Attach(A->cmap->N,1,*a->grid,0,0,z,A->cmap->n); 322e8146892SSatish Balay El::Gemv(El::TRANSPOSE,one,*a->emat,xe,one,ze); 323e883f9d5SXuan Zhou } 324e6dea9dbSXuan Zhou ierr = VecRestoreArrayRead(X,(const PetscScalar **)&x);CHKERRQ(ierr); 325e6dea9dbSXuan Zhou ierr = VecRestoreArray(Z,(PetscScalar **)&z);CHKERRQ(ierr); 326e883f9d5SXuan Zhou PetscFunctionReturn(0); 327e883f9d5SXuan Zhou } 328e883f9d5SXuan Zhou 3299a9e8502SHong Zhang static PetscErrorCode MatMatMultNumeric_Elemental(Mat A,Mat B,Mat C) 330c1d1b975SXuan Zhou { 331c1d1b975SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 332c1d1b975SXuan Zhou Mat_Elemental *b = (Mat_Elemental*)B->data; 3339a9e8502SHong Zhang Mat_Elemental *c = (Mat_Elemental*)C->data; 334e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 335c1d1b975SXuan Zhou 336c1d1b975SXuan Zhou PetscFunctionBegin; 337aae2c449SHong Zhang { /* Scoping so that constructor is called before pointer is returned */ 338e8146892SSatish Balay El::Gemm(El::NORMAL,El::NORMAL,one,*a->emat,*b->emat,zero,*c->emat); 339aae2c449SHong Zhang } 3409a9e8502SHong Zhang C->assembled = PETSC_TRUE; 3419a9e8502SHong Zhang PetscFunctionReturn(0); 3429a9e8502SHong Zhang } 3439a9e8502SHong Zhang 3449a9e8502SHong Zhang static PetscErrorCode MatMatMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C) 3459a9e8502SHong Zhang { 3469a9e8502SHong Zhang PetscErrorCode ierr; 3479a9e8502SHong Zhang Mat Ce; 348ce94432eSBarry Smith MPI_Comm comm; 3499a9e8502SHong Zhang 3509a9e8502SHong Zhang PetscFunctionBegin; 351ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 3529a9e8502SHong Zhang ierr = MatCreate(comm,&Ce);CHKERRQ(ierr); 3539a9e8502SHong Zhang ierr = MatSetSizes(Ce,A->rmap->n,B->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 3549a9e8502SHong Zhang ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr); 3559a9e8502SHong Zhang ierr = MatSetUp(Ce);CHKERRQ(ierr); 3569a9e8502SHong Zhang *C = Ce; 3579a9e8502SHong Zhang PetscFunctionReturn(0); 3589a9e8502SHong Zhang } 3599a9e8502SHong Zhang 3609a9e8502SHong Zhang static PetscErrorCode MatMatMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 3619a9e8502SHong Zhang { 3629a9e8502SHong Zhang PetscErrorCode ierr; 3639a9e8502SHong Zhang 3649a9e8502SHong Zhang PetscFunctionBegin; 3659a9e8502SHong Zhang if (scall == MAT_INITIAL_MATRIX){ 3663ff4c91cSHong Zhang ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 3679a9e8502SHong Zhang ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr); 3683ff4c91cSHong Zhang ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 3699a9e8502SHong Zhang } 3703ff4c91cSHong Zhang ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 3719a9e8502SHong Zhang ierr = MatMatMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr); 3723ff4c91cSHong Zhang ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 373c1d1b975SXuan Zhou PetscFunctionReturn(0); 374c1d1b975SXuan Zhou } 375c1d1b975SXuan Zhou 376df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultNumeric_Elemental(Mat A,Mat B,Mat C) 377df311e6cSXuan Zhou { 378df311e6cSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 379df311e6cSXuan Zhou Mat_Elemental *b = (Mat_Elemental*)B->data; 380df311e6cSXuan Zhou Mat_Elemental *c = (Mat_Elemental*)C->data; 381e6dea9dbSXuan Zhou PetscElemScalar one = 1,zero = 0; 382df311e6cSXuan Zhou 383df311e6cSXuan Zhou PetscFunctionBegin; 384df311e6cSXuan Zhou { /* Scoping so that constructor is called before pointer is returned */ 385e8146892SSatish Balay El::Gemm(El::NORMAL,El::TRANSPOSE,one,*a->emat,*b->emat,zero,*c->emat); 386df311e6cSXuan Zhou } 387df311e6cSXuan Zhou C->assembled = PETSC_TRUE; 388df311e6cSXuan Zhou PetscFunctionReturn(0); 389df311e6cSXuan Zhou } 390df311e6cSXuan Zhou 391df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMultSymbolic_Elemental(Mat A,Mat B,PetscReal fill,Mat *C) 392df311e6cSXuan Zhou { 393df311e6cSXuan Zhou PetscErrorCode ierr; 394df311e6cSXuan Zhou Mat Ce; 395ce94432eSBarry Smith MPI_Comm comm; 396df311e6cSXuan Zhou 397df311e6cSXuan Zhou PetscFunctionBegin; 398ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 399df311e6cSXuan Zhou ierr = MatCreate(comm,&Ce);CHKERRQ(ierr); 400df311e6cSXuan Zhou ierr = MatSetSizes(Ce,A->rmap->n,B->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 401df311e6cSXuan Zhou ierr = MatSetType(Ce,MATELEMENTAL);CHKERRQ(ierr); 402df311e6cSXuan Zhou ierr = MatSetUp(Ce);CHKERRQ(ierr); 403df311e6cSXuan Zhou *C = Ce; 404df311e6cSXuan Zhou PetscFunctionReturn(0); 405df311e6cSXuan Zhou } 406df311e6cSXuan Zhou 407df311e6cSXuan Zhou static PetscErrorCode MatMatTransposeMult_Elemental(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 408df311e6cSXuan Zhou { 409df311e6cSXuan Zhou PetscErrorCode ierr; 410df311e6cSXuan Zhou 411df311e6cSXuan Zhou PetscFunctionBegin; 412df311e6cSXuan Zhou if (scall == MAT_INITIAL_MATRIX){ 413df311e6cSXuan Zhou ierr = PetscLogEventBegin(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 414df311e6cSXuan Zhou ierr = MatMatMultSymbolic_Elemental(A,B,1.0,C);CHKERRQ(ierr); 415df311e6cSXuan Zhou ierr = PetscLogEventEnd(MAT_MatTransposeMultSymbolic,A,B,0,0);CHKERRQ(ierr); 416df311e6cSXuan Zhou } 417df311e6cSXuan Zhou ierr = PetscLogEventBegin(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 418df311e6cSXuan Zhou ierr = MatMatTransposeMultNumeric_Elemental(A,B,*C);CHKERRQ(ierr); 419df311e6cSXuan Zhou ierr = PetscLogEventEnd(MAT_MatTransposeMultNumeric,A,B,0,0);CHKERRQ(ierr); 420df311e6cSXuan Zhou PetscFunctionReturn(0); 421df311e6cSXuan Zhou } 422df311e6cSXuan Zhou 423a9d89745SXuan Zhou static PetscErrorCode MatGetDiagonal_Elemental(Mat A,Vec D) 42461119200SXuan Zhou { 425a9d89745SXuan Zhou PetscInt i,nrows,ncols,nD,rrank,ridx,crank,cidx; 426a9d89745SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 42761119200SXuan Zhou PetscErrorCode ierr; 428a9d89745SXuan Zhou PetscElemScalar v; 429ce94432eSBarry Smith MPI_Comm comm; 43061119200SXuan Zhou 43161119200SXuan Zhou PetscFunctionBegin; 432ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 433a9d89745SXuan Zhou ierr = MatGetSize(A,&nrows,&ncols);CHKERRQ(ierr); 434a9d89745SXuan Zhou nD = nrows>ncols ? ncols : nrows; 435a9d89745SXuan Zhou for (i=0; i<nD; i++) { 436a9d89745SXuan Zhou PetscInt erow,ecol; 437a9d89745SXuan Zhou P2RO(A,0,i,&rrank,&ridx); 438a9d89745SXuan Zhou RO2E(A,0,rrank,ridx,&erow); 439a9d89745SXuan Zhou if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation"); 440a9d89745SXuan Zhou P2RO(A,1,i,&crank,&cidx); 441a9d89745SXuan Zhou RO2E(A,1,crank,cidx,&ecol); 442a9d89745SXuan Zhou if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation"); 443a9d89745SXuan Zhou v = a->emat->Get(erow,ecol); 4447c8b904fSSatish Balay ierr = VecSetValues(D,1,&i,(PetscScalar*)&v,INSERT_VALUES);CHKERRQ(ierr); 445ade3cc5eSXuan Zhou } 446a9d89745SXuan Zhou ierr = VecAssemblyBegin(D);CHKERRQ(ierr); 447a9d89745SXuan Zhou ierr = VecAssemblyEnd(D);CHKERRQ(ierr); 44861119200SXuan Zhou PetscFunctionReturn(0); 44961119200SXuan Zhou } 45061119200SXuan Zhou 451ade3cc5eSXuan Zhou static PetscErrorCode MatDiagonalScale_Elemental(Mat X,Vec L,Vec R) 452ade3cc5eSXuan Zhou { 453ade3cc5eSXuan Zhou Mat_Elemental *x = (Mat_Elemental*)X->data; 454ade3cc5eSXuan Zhou const PetscElemScalar *d; 455ade3cc5eSXuan Zhou PetscErrorCode ierr; 456ade3cc5eSXuan Zhou 457ade3cc5eSXuan Zhou PetscFunctionBegin; 4589065cd98SJed Brown if (R) { 459ade3cc5eSXuan Zhou ierr = VecGetArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr); 460e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de; 4610c18141cSBarry Smith de.LockedAttach(X->cmap->N,1,*x->grid,0,0,d,X->cmap->n); 462e8146892SSatish Balay El::DiagonalScale(El::RIGHT,El::NORMAL,de,*x->emat); 463ade3cc5eSXuan Zhou ierr = VecRestoreArrayRead(R,(const PetscScalar **)&d);CHKERRQ(ierr); 4649065cd98SJed Brown } 4659065cd98SJed Brown if (L) { 466ade3cc5eSXuan Zhou ierr = VecGetArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr); 467e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> de; 4680c18141cSBarry Smith de.LockedAttach(X->rmap->N,1,*x->grid,0,0,d,X->rmap->n); 469e8146892SSatish Balay El::DiagonalScale(El::LEFT,El::NORMAL,de,*x->emat); 470ade3cc5eSXuan Zhou ierr = VecRestoreArrayRead(L,(const PetscScalar **)&d);CHKERRQ(ierr); 471ade3cc5eSXuan Zhou } 472ade3cc5eSXuan Zhou PetscFunctionReturn(0); 473ade3cc5eSXuan Zhou } 474ade3cc5eSXuan Zhou 475e6dea9dbSXuan Zhou static PetscErrorCode MatScale_Elemental(Mat X,PetscScalar a) 47665b78793SXuan Zhou { 47765b78793SXuan Zhou Mat_Elemental *x = (Mat_Elemental*)X->data; 47865b78793SXuan Zhou 47965b78793SXuan Zhou PetscFunctionBegin; 480e8146892SSatish Balay El::Scale((PetscElemScalar)a,*x->emat); 48165b78793SXuan Zhou PetscFunctionReturn(0); 48265b78793SXuan Zhou } 48365b78793SXuan Zhou 484f82baa17SHong Zhang /* 485f82baa17SHong Zhang MatAXPY - Computes Y = a*X + Y. 486f82baa17SHong Zhang */ 487e6dea9dbSXuan Zhou static PetscErrorCode MatAXPY_Elemental(Mat Y,PetscScalar a,Mat X,MatStructure str) 488e09a3074SHong Zhang { 489e09a3074SHong Zhang Mat_Elemental *x = (Mat_Elemental*)X->data; 490e09a3074SHong Zhang Mat_Elemental *y = (Mat_Elemental*)Y->data; 49168446cf8SJed Brown PetscErrorCode ierr; 492e09a3074SHong Zhang 493e09a3074SHong Zhang PetscFunctionBegin; 494e8146892SSatish Balay El::Axpy((PetscElemScalar)a,*x->emat,*y->emat); 49521f6c9c4SJed Brown ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 496e09a3074SHong Zhang PetscFunctionReturn(0); 497e09a3074SHong Zhang } 498e09a3074SHong Zhang 499d6223691SXuan Zhou static PetscErrorCode MatCopy_Elemental(Mat A,Mat B,MatStructure str) 500d6223691SXuan Zhou { 501d6223691SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 502d6223691SXuan Zhou Mat_Elemental *b=(Mat_Elemental*)B->data; 503*a4ee3bb6SSatish Balay PetscErrorCode ierr; 504d6223691SXuan Zhou 505d6223691SXuan Zhou PetscFunctionBegin; 506e8146892SSatish Balay El::Copy(*a->emat,*b->emat); 507cdc753b6SBarry Smith ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 508d6223691SXuan Zhou PetscFunctionReturn(0); 509d6223691SXuan Zhou } 510d6223691SXuan Zhou 511df311e6cSXuan Zhou static PetscErrorCode MatDuplicate_Elemental(Mat A,MatDuplicateOption op,Mat *B) 512df311e6cSXuan Zhou { 513df311e6cSXuan Zhou Mat Be; 514ce94432eSBarry Smith MPI_Comm comm; 515df311e6cSXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 516df311e6cSXuan Zhou PetscErrorCode ierr; 517df311e6cSXuan Zhou 518df311e6cSXuan Zhou PetscFunctionBegin; 519ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 520df311e6cSXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 521df311e6cSXuan Zhou ierr = MatSetSizes(Be,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 522df311e6cSXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 523df311e6cSXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 524df311e6cSXuan Zhou *B = Be; 525df311e6cSXuan Zhou if (op == MAT_COPY_VALUES) { 526df311e6cSXuan Zhou Mat_Elemental *b=(Mat_Elemental*)Be->data; 527e8146892SSatish Balay El::Copy(*a->emat,*b->emat); 528df311e6cSXuan Zhou } 529df311e6cSXuan Zhou Be->assembled = PETSC_TRUE; 530df311e6cSXuan Zhou PetscFunctionReturn(0); 531df311e6cSXuan Zhou } 532df311e6cSXuan Zhou 533d6223691SXuan Zhou static PetscErrorCode MatTranspose_Elemental(Mat A,MatReuse reuse,Mat *B) 534d6223691SXuan Zhou { 535ec8cb81fSBarry Smith Mat Be = *B; 5365262d616SXuan Zhou PetscErrorCode ierr; 537ce94432eSBarry Smith MPI_Comm comm; 5384fe7bbcaSHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data, *b; 539d6223691SXuan Zhou 540d6223691SXuan Zhou PetscFunctionBegin; 541ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5425cb544a0SHong Zhang /* Only out-of-place supported */ 5435262d616SXuan Zhou if (reuse == MAT_INITIAL_MATRIX){ 5445262d616SXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 5455262d616SXuan Zhou ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 5465262d616SXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 5475262d616SXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 5485262d616SXuan Zhou *B = Be; 5495262d616SXuan Zhou } 5504fe7bbcaSHong Zhang b = (Mat_Elemental*)Be->data; 551e8146892SSatish Balay El::Transpose(*a->emat,*b->emat); 5525262d616SXuan Zhou Be->assembled = PETSC_TRUE; 553d6223691SXuan Zhou PetscFunctionReturn(0); 554d6223691SXuan Zhou } 555d6223691SXuan Zhou 556dfcb0403SXuan Zhou static PetscErrorCode MatConjugate_Elemental(Mat A) 557dfcb0403SXuan Zhou { 558dfcb0403SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 559dfcb0403SXuan Zhou 560dfcb0403SXuan Zhou PetscFunctionBegin; 561e8146892SSatish Balay El::Conjugate(*a->emat); 562dfcb0403SXuan Zhou PetscFunctionReturn(0); 563dfcb0403SXuan Zhou } 564dfcb0403SXuan Zhou 5654a29722dSXuan Zhou static PetscErrorCode MatHermitianTranspose_Elemental(Mat A,MatReuse reuse,Mat *B) 5664a29722dSXuan Zhou { 567ec8cb81fSBarry Smith Mat Be = *B; 5684a29722dSXuan Zhou PetscErrorCode ierr; 569ce94432eSBarry Smith MPI_Comm comm; 5704a29722dSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data, *b; 5714a29722dSXuan Zhou 5724a29722dSXuan Zhou PetscFunctionBegin; 573ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5745cb544a0SHong Zhang /* Only out-of-place supported */ 5754a29722dSXuan Zhou if (reuse == MAT_INITIAL_MATRIX){ 5764a29722dSXuan Zhou ierr = MatCreate(comm,&Be);CHKERRQ(ierr); 5774a29722dSXuan Zhou ierr = MatSetSizes(Be,A->cmap->n,A->rmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 5784a29722dSXuan Zhou ierr = MatSetType(Be,MATELEMENTAL);CHKERRQ(ierr); 5794a29722dSXuan Zhou ierr = MatSetUp(Be);CHKERRQ(ierr); 5804a29722dSXuan Zhou *B = Be; 5814a29722dSXuan Zhou } 5824a29722dSXuan Zhou b = (Mat_Elemental*)Be->data; 583e8146892SSatish Balay El::Adjoint(*a->emat,*b->emat); 5844a29722dSXuan Zhou Be->assembled = PETSC_TRUE; 5854a29722dSXuan Zhou PetscFunctionReturn(0); 5864a29722dSXuan Zhou } 5874a29722dSXuan Zhou 5881f881ff8SXuan Zhou static PetscErrorCode MatSolve_Elemental(Mat A,Vec B,Vec X) 5891f881ff8SXuan Zhou { 5901f881ff8SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 5911f881ff8SXuan Zhou PetscErrorCode ierr; 592df311e6cSXuan Zhou PetscElemScalar *x; 593ea394475SSatish Balay PetscInt pivoting = a->pivoting; 5941f881ff8SXuan Zhou 5951f881ff8SXuan Zhou PetscFunctionBegin; 59645cf121fSXuan Zhou ierr = VecCopy(B,X);CHKERRQ(ierr); 597e6dea9dbSXuan Zhou ierr = VecGetArray(X,(PetscScalar **)&x);CHKERRQ(ierr); 598ea394475SSatish Balay 599e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::VC,El::STAR> xe; 6000c18141cSBarry Smith xe.Attach(A->rmap->N,1,*a->grid,0,0,x,A->rmap->n); 601e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::MC,El::MR> xer(xe); 602fc54b460SXuan Zhou switch (A->factortype) { 603fc54b460SXuan Zhou case MAT_FACTOR_LU: 604ea394475SSatish Balay if (pivoting == 0) { 605e8146892SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,xer); 606ea394475SSatish Balay } else if (pivoting == 1) { 607ea394475SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,xer); 608ea394475SSatish Balay } else { /* pivoting == 2 */ 609ea394475SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,xer); 6101f881ff8SXuan Zhou } 611fc54b460SXuan Zhou break; 612fc54b460SXuan Zhou case MAT_FACTOR_CHOLESKY: 613e8146892SSatish Balay El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,xer); 614fc54b460SXuan Zhou break; 615fc54b460SXuan Zhou default: 6164fe7bbcaSHong Zhang SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType"); 617fc54b460SXuan Zhou break; 6181f881ff8SXuan Zhou } 619ea394475SSatish Balay El::Copy(xer,xe); 620ea394475SSatish Balay 621e6dea9dbSXuan Zhou ierr = VecRestoreArray(X,(PetscScalar **)&x);CHKERRQ(ierr); 6221f881ff8SXuan Zhou PetscFunctionReturn(0); 6231f881ff8SXuan Zhou } 6241f881ff8SXuan Zhou 625df311e6cSXuan Zhou static PetscErrorCode MatSolveAdd_Elemental(Mat A,Vec B,Vec Y,Vec X) 626df311e6cSXuan Zhou { 627df311e6cSXuan Zhou PetscErrorCode ierr; 628df311e6cSXuan Zhou 629df311e6cSXuan Zhou PetscFunctionBegin; 630df311e6cSXuan Zhou ierr = MatSolve_Elemental(A,B,X);CHKERRQ(ierr); 6313d7f40dbSXuan Zhou ierr = VecAXPY(X,1,Y);CHKERRQ(ierr); 632df311e6cSXuan Zhou PetscFunctionReturn(0); 633df311e6cSXuan Zhou } 634df311e6cSXuan Zhou 635ae844d54SHong Zhang static PetscErrorCode MatMatSolve_Elemental(Mat A,Mat B,Mat X) 636ae844d54SHong Zhang { 6371f0e42cfSHong Zhang Mat_Elemental *a=(Mat_Elemental*)A->data; 638d6223691SXuan Zhou Mat_Elemental *b=(Mat_Elemental*)B->data; 6391f0e42cfSHong Zhang Mat_Elemental *x=(Mat_Elemental*)X->data; 640ea394475SSatish Balay PetscInt pivoting = a->pivoting; 6411f0e42cfSHong Zhang 642ae844d54SHong Zhang PetscFunctionBegin; 643e8146892SSatish Balay El::Copy(*b->emat,*x->emat); 644fc54b460SXuan Zhou switch (A->factortype) { 645fc54b460SXuan Zhou case MAT_FACTOR_LU: 646ea394475SSatish Balay if (pivoting == 0) { 647e8146892SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,*x->emat); 648ea394475SSatish Balay } else if (pivoting == 1) { 649ea394475SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*x->emat); 650ea394475SSatish Balay } else { 651ea394475SSatish Balay El::lu::SolveAfter(El::NORMAL,*a->emat,*a->P,*a->Q,*x->emat); 652d6223691SXuan Zhou } 653fc54b460SXuan Zhou break; 654fc54b460SXuan Zhou case MAT_FACTOR_CHOLESKY: 655e8146892SSatish Balay El::cholesky::SolveAfter(El::UPPER,El::NORMAL,*a->emat,*x->emat); 656fc54b460SXuan Zhou break; 657fc54b460SXuan Zhou default: 6584fe7bbcaSHong Zhang SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unfactored Matrix or Unsupported MatFactorType"); 659fc54b460SXuan Zhou break; 660fc54b460SXuan Zhou } 661ae844d54SHong Zhang PetscFunctionReturn(0); 662ae844d54SHong Zhang } 663ae844d54SHong Zhang 664ae844d54SHong Zhang static PetscErrorCode MatLUFactor_Elemental(Mat A,IS row,IS col,const MatFactorInfo *info) 665ae844d54SHong Zhang { 6667c920d81SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 667124af5d2SSatish Balay PetscErrorCode ierr; 668ea394475SSatish Balay PetscInt pivoting = a->pivoting; 6697c920d81SXuan Zhou 670ae844d54SHong Zhang PetscFunctionBegin; 671ea394475SSatish Balay if (pivoting == 0) { 672e8146892SSatish Balay El::LU(*a->emat); 673ea394475SSatish Balay } else if (pivoting == 1) { 674ea394475SSatish Balay El::LU(*a->emat,*a->P); 675ea394475SSatish Balay } else { 676ea394475SSatish Balay El::LU(*a->emat,*a->P,*a->Q); 677d6223691SXuan Zhou } 6781293973dSHong Zhang A->factortype = MAT_FACTOR_LU; 679834d3fecSHong Zhang A->assembled = PETSC_TRUE; 680f6224b95SHong Zhang 681f6224b95SHong Zhang ierr = PetscFree(A->solvertype);CHKERRQ(ierr); 682f6224b95SHong Zhang ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);CHKERRQ(ierr); 683ae844d54SHong Zhang PetscFunctionReturn(0); 684ae844d54SHong Zhang } 685ae844d54SHong Zhang 686d7c3f9d8SHong Zhang static PetscErrorCode MatLUFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info) 687d7c3f9d8SHong Zhang { 688d7c3f9d8SHong Zhang PetscErrorCode ierr; 689d7c3f9d8SHong Zhang 690d7c3f9d8SHong Zhang PetscFunctionBegin; 691d7c3f9d8SHong Zhang ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 692d7c3f9d8SHong Zhang ierr = MatLUFactor_Elemental(F,0,0,info);CHKERRQ(ierr); 693d7c3f9d8SHong Zhang PetscFunctionReturn(0); 694d7c3f9d8SHong Zhang } 695d7c3f9d8SHong Zhang 696d7c3f9d8SHong Zhang static PetscErrorCode MatLUFactorSymbolic_Elemental(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 697d7c3f9d8SHong Zhang { 698d7c3f9d8SHong Zhang PetscFunctionBegin; 699d7c3f9d8SHong Zhang /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */ 700d7c3f9d8SHong Zhang PetscFunctionReturn(0); 701d7c3f9d8SHong Zhang } 702d7c3f9d8SHong Zhang 70345cf121fSXuan Zhou static PetscErrorCode MatCholeskyFactor_Elemental(Mat A,IS perm,const MatFactorInfo *info) 70445cf121fSXuan Zhou { 70545cf121fSXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 706e8146892SSatish Balay El::DistMatrix<PetscElemScalar,El::MC,El::STAR> d; 707124af5d2SSatish Balay PetscErrorCode ierr; 70845cf121fSXuan Zhou 70945cf121fSXuan Zhou PetscFunctionBegin; 710e8146892SSatish Balay El::Cholesky(El::UPPER,*a->emat); 711fc54b460SXuan Zhou A->factortype = MAT_FACTOR_CHOLESKY; 712834d3fecSHong Zhang A->assembled = PETSC_TRUE; 713f6224b95SHong Zhang 714f6224b95SHong Zhang ierr = PetscFree(A->solvertype);CHKERRQ(ierr); 715f6224b95SHong Zhang ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&A->solvertype);CHKERRQ(ierr); 71645cf121fSXuan Zhou PetscFunctionReturn(0); 71745cf121fSXuan Zhou } 71845cf121fSXuan Zhou 71979673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorNumeric_Elemental(Mat F,Mat A,const MatFactorInfo *info) 72079673f7bSHong Zhang { 721cb76c1d8SXuan Zhou PetscErrorCode ierr; 722cb76c1d8SXuan Zhou 723cb76c1d8SXuan Zhou PetscFunctionBegin; 724cb76c1d8SXuan Zhou ierr = MatCopy(A,F,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 725cb76c1d8SXuan Zhou ierr = MatCholeskyFactor_Elemental(F,0,info);CHKERRQ(ierr); 726cb76c1d8SXuan Zhou PetscFunctionReturn(0); 72779673f7bSHong Zhang } 728cb76c1d8SXuan Zhou 72979673f7bSHong Zhang static PetscErrorCode MatCholeskyFactorSymbolic_Elemental(Mat F,Mat A,IS perm,const MatFactorInfo *info) 73079673f7bSHong Zhang { 73179673f7bSHong Zhang PetscFunctionBegin; 73279673f7bSHong Zhang /* F is create and allocated by MatGetFactor_elemental_petsc(), skip this routine. */ 73379673f7bSHong Zhang PetscFunctionReturn(0); 73479673f7bSHong Zhang } 73579673f7bSHong Zhang 73615767789SHong Zhang PetscErrorCode MatFactorGetSolverPackage_elemental_elemental(Mat A,const MatSolverPackage *type) 73715767789SHong Zhang { 73815767789SHong Zhang PetscFunctionBegin; 73915767789SHong Zhang *type = MATSOLVERELEMENTAL; 74015767789SHong Zhang PetscFunctionReturn(0); 74115767789SHong Zhang } 74215767789SHong Zhang 74315767789SHong Zhang static PetscErrorCode MatGetFactor_elemental_elemental(Mat A,MatFactorType ftype,Mat *F) 7441293973dSHong Zhang { 7451293973dSHong Zhang Mat B; 7461293973dSHong Zhang PetscErrorCode ierr; 7471293973dSHong Zhang 7481293973dSHong Zhang PetscFunctionBegin; 7491293973dSHong Zhang /* Create the factorization matrix */ 750ce94432eSBarry Smith ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 7511293973dSHong Zhang ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 7521293973dSHong Zhang ierr = MatSetType(B,MATELEMENTAL);CHKERRQ(ierr); 7531293973dSHong Zhang ierr = MatSetUp(B);CHKERRQ(ierr); 7541293973dSHong Zhang B->factortype = ftype; 75500c67f3bSHong Zhang ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 75600c67f3bSHong Zhang ierr = PetscStrallocpy(MATSOLVERELEMENTAL,&B->solvertype);CHKERRQ(ierr); 75700c67f3bSHong Zhang 758bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_elemental_elemental);CHKERRQ(ierr); 7591293973dSHong Zhang *F = B; 7601293973dSHong Zhang PetscFunctionReturn(0); 7611293973dSHong Zhang } 7621293973dSHong Zhang 76329b38603SBarry Smith PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_Elemental(void) 76442c9c57cSBarry Smith { 76518713533SBarry Smith PetscErrorCode ierr; 76618713533SBarry Smith 76742c9c57cSBarry Smith PetscFunctionBegin; 76842c9c57cSBarry Smith ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL, MAT_FACTOR_LU,MatGetFactor_elemental_elemental);CHKERRQ(ierr); 76942c9c57cSBarry Smith ierr = MatSolverPackageRegister(MATSOLVERELEMENTAL,MATELEMENTAL, MAT_FACTOR_CHOLESKY,MatGetFactor_elemental_elemental);CHKERRQ(ierr); 77042c9c57cSBarry Smith PetscFunctionReturn(0); 77142c9c57cSBarry Smith } 77242c9c57cSBarry Smith 7731f881ff8SXuan Zhou static PetscErrorCode MatNorm_Elemental(Mat A,NormType type,PetscReal *nrm) 7741f881ff8SXuan Zhou { 7751f881ff8SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 7761f881ff8SXuan Zhou 7771f881ff8SXuan Zhou PetscFunctionBegin; 7781f881ff8SXuan Zhou switch (type){ 7791f881ff8SXuan Zhou case NORM_1: 780e8146892SSatish Balay *nrm = El::OneNorm(*a->emat); 7811f881ff8SXuan Zhou break; 7821f881ff8SXuan Zhou case NORM_FROBENIUS: 783e8146892SSatish Balay *nrm = El::FrobeniusNorm(*a->emat); 7841f881ff8SXuan Zhou break; 7851f881ff8SXuan Zhou case NORM_INFINITY: 786e8146892SSatish Balay *nrm = El::InfinityNorm(*a->emat); 7871f881ff8SXuan Zhou break; 7881f881ff8SXuan Zhou default: 7891f881ff8SXuan Zhou printf("Error: unsupported norm type!\n"); 7901f881ff8SXuan Zhou } 7911f881ff8SXuan Zhou PetscFunctionReturn(0); 7921f881ff8SXuan Zhou } 7931f881ff8SXuan Zhou 7945262d616SXuan Zhou static PetscErrorCode MatZeroEntries_Elemental(Mat A) 7955262d616SXuan Zhou { 7965262d616SXuan Zhou Mat_Elemental *a=(Mat_Elemental*)A->data; 7975262d616SXuan Zhou 7985262d616SXuan Zhou PetscFunctionBegin; 799e8146892SSatish Balay El::Zero(*a->emat); 8005262d616SXuan Zhou PetscFunctionReturn(0); 8015262d616SXuan Zhou } 8025262d616SXuan Zhou 803db31f6deSJed Brown static PetscErrorCode MatGetOwnershipIS_Elemental(Mat A,IS *rows,IS *cols) 804db31f6deSJed Brown { 805db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 806db31f6deSJed Brown PetscErrorCode ierr; 807db31f6deSJed Brown PetscInt i,m,shift,stride,*idx; 808db31f6deSJed Brown 809db31f6deSJed Brown PetscFunctionBegin; 810db31f6deSJed Brown if (rows) { 811db31f6deSJed Brown m = a->emat->LocalHeight(); 812db31f6deSJed Brown shift = a->emat->ColShift(); 813db31f6deSJed Brown stride = a->emat->ColStride(); 814785e854fSJed Brown ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 815db31f6deSJed Brown for (i=0; i<m; i++) { 816db31f6deSJed Brown PetscInt rank,offset; 817db31f6deSJed Brown E2RO(A,0,shift+i*stride,&rank,&offset); 818db31f6deSJed Brown RO2P(A,0,rank,offset,&idx[i]); 819db31f6deSJed Brown } 820db31f6deSJed Brown ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,rows);CHKERRQ(ierr); 821db31f6deSJed Brown } 822db31f6deSJed Brown if (cols) { 823db31f6deSJed Brown m = a->emat->LocalWidth(); 824db31f6deSJed Brown shift = a->emat->RowShift(); 825db31f6deSJed Brown stride = a->emat->RowStride(); 826785e854fSJed Brown ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 827db31f6deSJed Brown for (i=0; i<m; i++) { 828db31f6deSJed Brown PetscInt rank,offset; 829db31f6deSJed Brown E2RO(A,1,shift+i*stride,&rank,&offset); 830db31f6deSJed Brown RO2P(A,1,rank,offset,&idx[i]); 831db31f6deSJed Brown } 832db31f6deSJed Brown ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,cols);CHKERRQ(ierr); 833db31f6deSJed Brown } 834db31f6deSJed Brown PetscFunctionReturn(0); 835db31f6deSJed Brown } 836db31f6deSJed Brown 83719fd82e9SBarry Smith static PetscErrorCode MatConvert_Elemental_Dense(Mat A,MatType newtype,MatReuse reuse,Mat *B) 838af295397SXuan Zhou { 8392ef0cf24SXuan Zhou Mat Bmpi; 840af295397SXuan Zhou Mat_Elemental *a = (Mat_Elemental*)A->data; 841ce94432eSBarry Smith MPI_Comm comm; 8422ef0cf24SXuan Zhou PetscErrorCode ierr; 843fa9e26c8SHong Zhang IS isrows,iscols; 844fa9e26c8SHong Zhang PetscInt rrank,ridx,crank,cidx,nrows,ncols,i,j,erow,ecol,elrow,elcol; 845fa9e26c8SHong Zhang const PetscInt *rows,*cols; 846df311e6cSXuan Zhou PetscElemScalar v; 847fa9e26c8SHong Zhang const El::Grid &grid = a->emat->Grid(); 848af295397SXuan Zhou 849af295397SXuan Zhou PetscFunctionBegin; 850ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 851a000e53bSHong Zhang 852de0a22f0SHong Zhang if (reuse == MAT_REUSE_MATRIX) { 853de0a22f0SHong Zhang Bmpi = *B; 854de0a22f0SHong Zhang } else { 855af295397SXuan Zhou ierr = MatCreate(comm,&Bmpi);CHKERRQ(ierr); 856af295397SXuan Zhou ierr = MatSetSizes(Bmpi,A->rmap->n,A->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 8572ef0cf24SXuan Zhou ierr = MatSetType(Bmpi,MATDENSE);CHKERRQ(ierr); 858af295397SXuan Zhou ierr = MatSetUp(Bmpi);CHKERRQ(ierr); 859de0a22f0SHong Zhang } 860fa9e26c8SHong Zhang 861fa9e26c8SHong Zhang /* Get local entries of A */ 862fa9e26c8SHong Zhang ierr = MatGetOwnershipIS(A,&isrows,&iscols);CHKERRQ(ierr); 863fa9e26c8SHong Zhang ierr = ISGetLocalSize(isrows,&nrows);CHKERRQ(ierr); 864fa9e26c8SHong Zhang ierr = ISGetIndices(isrows,&rows);CHKERRQ(ierr); 865fa9e26c8SHong Zhang ierr = ISGetLocalSize(iscols,&ncols);CHKERRQ(ierr); 866fa9e26c8SHong Zhang ierr = ISGetIndices(iscols,&cols);CHKERRQ(ierr); 867fa9e26c8SHong Zhang 86857299f2fSHong Zhang if (a->roworiented) { 8692ef0cf24SXuan Zhou for (i=0; i<nrows; i++) { 870fa9e26c8SHong Zhang P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */ 8712ef0cf24SXuan Zhou RO2E(A,0,rrank,ridx,&erow); 8722ef0cf24SXuan Zhou if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation"); 8732ef0cf24SXuan Zhou for (j=0; j<ncols; j++) { 874fa9e26c8SHong Zhang P2RO(A,1,cols[j],&crank,&cidx); 8752ef0cf24SXuan Zhou RO2E(A,1,crank,cidx,&ecol); 8762ef0cf24SXuan Zhou if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation"); 877fa9e26c8SHong Zhang 878fa9e26c8SHong Zhang elrow = erow / grid.MCSize(); /* Elemental local row index */ 879fa9e26c8SHong Zhang elcol = ecol / grid.MRSize(); /* Elemental local column index */ 880fa9e26c8SHong Zhang v = a->emat->GetLocal(elrow,elcol); 881fa9e26c8SHong Zhang ierr = MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr); 8822ef0cf24SXuan Zhou } 8832ef0cf24SXuan Zhou } 88457299f2fSHong Zhang } else { /* column-oriented */ 88557299f2fSHong Zhang for (j=0; j<ncols; j++) { 88657299f2fSHong Zhang P2RO(A,1,cols[j],&crank,&cidx); 88757299f2fSHong Zhang RO2E(A,1,crank,cidx,&ecol); 88857299f2fSHong Zhang if (crank < 0 || cidx < 0 || ecol < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect col translation"); 88957299f2fSHong Zhang for (i=0; i<nrows; i++) { 89057299f2fSHong Zhang P2RO(A,0,rows[i],&rrank,&ridx); /* convert indices between PETSc <-> (Rank,Offset) <-> Elemental */ 89157299f2fSHong Zhang RO2E(A,0,rrank,ridx,&erow); 89257299f2fSHong Zhang if (rrank < 0 || ridx < 0 || erow < 0) SETERRQ(comm,PETSC_ERR_PLIB,"Incorrect row translation"); 89357299f2fSHong Zhang 89457299f2fSHong Zhang elrow = erow / grid.MCSize(); /* Elemental local row index */ 89557299f2fSHong Zhang elcol = ecol / grid.MRSize(); /* Elemental local column index */ 89657299f2fSHong Zhang v = a->emat->GetLocal(elrow,elcol); 89757299f2fSHong Zhang ierr = MatSetValues(Bmpi,1,&rows[i],1,&cols[j],(PetscScalar *)&v,INSERT_VALUES);CHKERRQ(ierr); 89857299f2fSHong Zhang } 89957299f2fSHong Zhang } 90057299f2fSHong Zhang } 901af295397SXuan Zhou ierr = MatAssemblyBegin(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 902af295397SXuan Zhou ierr = MatAssemblyEnd(Bmpi,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 903511c6705SHong Zhang if (reuse == MAT_INPLACE_MATRIX) { 90428be2f97SBarry Smith ierr = MatHeaderReplace(A,&Bmpi);CHKERRQ(ierr); 905c4ad791aSXuan Zhou } else { 906c4ad791aSXuan Zhou *B = Bmpi; 907c4ad791aSXuan Zhou } 908fa9e26c8SHong Zhang ierr = ISDestroy(&isrows);CHKERRQ(ierr); 909fa9e26c8SHong Zhang ierr = ISDestroy(&iscols);CHKERRQ(ierr); 910af295397SXuan Zhou PetscFunctionReturn(0); 911af295397SXuan Zhou } 912af295397SXuan Zhou 913cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat) 914af8000cdSHong Zhang { 915af8000cdSHong Zhang Mat mat_elemental; 916af8000cdSHong Zhang PetscErrorCode ierr; 917af8000cdSHong Zhang PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols; 918af8000cdSHong Zhang const PetscInt *cols; 919af8000cdSHong Zhang const PetscScalar *vals; 920af8000cdSHong Zhang 921af8000cdSHong Zhang PetscFunctionBegin; 922af8000cdSHong Zhang ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr); 923af8000cdSHong Zhang ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 924af8000cdSHong Zhang ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr); 925af8000cdSHong Zhang ierr = MatSetUp(mat_elemental);CHKERRQ(ierr); 926af8000cdSHong Zhang for (row=0; row<M; row++) { 927af8000cdSHong Zhang ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 928af8000cdSHong Zhang /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */ 929af8000cdSHong Zhang ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 930af8000cdSHong Zhang ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 931af8000cdSHong Zhang } 932af8000cdSHong Zhang ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 933af8000cdSHong Zhang ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 934af8000cdSHong Zhang 935511c6705SHong Zhang if (reuse == MAT_INPLACE_MATRIX) { 93628be2f97SBarry Smith ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr); 937af8000cdSHong Zhang } else { 938af8000cdSHong Zhang *newmat = mat_elemental; 939af8000cdSHong Zhang } 940af8000cdSHong Zhang PetscFunctionReturn(0); 941af8000cdSHong Zhang } 942af8000cdSHong Zhang 943cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat) 9445d7652ecSHong Zhang { 9455d7652ecSHong Zhang Mat mat_elemental; 9465d7652ecSHong Zhang PetscErrorCode ierr; 9475d7652ecSHong Zhang PetscInt row,ncols,rstart=A->rmap->rstart,rend=A->rmap->rend,j; 9485d7652ecSHong Zhang const PetscInt *cols; 9495d7652ecSHong Zhang const PetscScalar *vals; 9505d7652ecSHong Zhang 9515d7652ecSHong Zhang PetscFunctionBegin; 9525d7652ecSHong Zhang ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr); 9535d7652ecSHong Zhang ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 9545d7652ecSHong Zhang ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr); 9555d7652ecSHong Zhang ierr = MatSetUp(mat_elemental);CHKERRQ(ierr); 9565d7652ecSHong Zhang for (row=rstart; row<rend; row++) { 9575d7652ecSHong Zhang ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 9585d7652ecSHong Zhang for (j=0; j<ncols; j++) { 9595d7652ecSHong Zhang /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */ 9605d7652ecSHong Zhang ierr = MatSetValues(mat_elemental,1,&row,1,&cols[j],&vals[j],ADD_VALUES);CHKERRQ(ierr); 9615d7652ecSHong Zhang } 9625d7652ecSHong Zhang ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 9635d7652ecSHong Zhang } 9645d7652ecSHong Zhang ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9655d7652ecSHong Zhang ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9665d7652ecSHong Zhang 967511c6705SHong Zhang if (reuse == MAT_INPLACE_MATRIX) { 96828be2f97SBarry Smith ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr); 9695d7652ecSHong Zhang } else { 9705d7652ecSHong Zhang *newmat = mat_elemental; 9715d7652ecSHong Zhang } 9725d7652ecSHong Zhang PetscFunctionReturn(0); 9735d7652ecSHong Zhang } 9745d7652ecSHong Zhang 975cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat) 9766214f412SHong Zhang { 9776214f412SHong Zhang Mat mat_elemental; 9786214f412SHong Zhang PetscErrorCode ierr; 9796214f412SHong Zhang PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols,j; 9806214f412SHong Zhang const PetscInt *cols; 9816214f412SHong Zhang const PetscScalar *vals; 9826214f412SHong Zhang 9836214f412SHong Zhang PetscFunctionBegin; 9846214f412SHong Zhang ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr); 9856214f412SHong Zhang ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 9866214f412SHong Zhang ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr); 9876214f412SHong Zhang ierr = MatSetUp(mat_elemental);CHKERRQ(ierr); 9886214f412SHong Zhang ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr); 9896214f412SHong Zhang for (row=0; row<M; row++) { 9906214f412SHong Zhang ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 9916214f412SHong Zhang /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */ 9926214f412SHong Zhang ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 9936214f412SHong Zhang for (j=0; j<ncols; j++) { /* lower triangular part */ 9946214f412SHong Zhang if (cols[j] == row) continue; 9956214f412SHong Zhang ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr); 9966214f412SHong Zhang } 9976214f412SHong Zhang ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 9986214f412SHong Zhang } 9996214f412SHong Zhang ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr); 10006214f412SHong Zhang ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10016214f412SHong Zhang ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10026214f412SHong Zhang 1003511c6705SHong Zhang if (reuse == MAT_INPLACE_MATRIX) { 100428be2f97SBarry Smith ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr); 10056214f412SHong Zhang } else { 10066214f412SHong Zhang *newmat = mat_elemental; 10076214f412SHong Zhang } 10086214f412SHong Zhang PetscFunctionReturn(0); 10096214f412SHong Zhang } 10106214f412SHong Zhang 1011cc2e6a90SBarry Smith PETSC_INTERN PetscErrorCode MatConvert_MPISBAIJ_Elemental(Mat A, MatType newtype,MatReuse reuse,Mat *newmat) 10126214f412SHong Zhang { 10136214f412SHong Zhang Mat mat_elemental; 10146214f412SHong Zhang PetscErrorCode ierr; 10156214f412SHong Zhang PetscInt M=A->rmap->N,N=A->cmap->N,row,ncols,j,rstart=A->rmap->rstart,rend=A->rmap->rend; 10166214f412SHong Zhang const PetscInt *cols; 10176214f412SHong Zhang const PetscScalar *vals; 10186214f412SHong Zhang 10196214f412SHong Zhang PetscFunctionBegin; 10206214f412SHong Zhang ierr = MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental);CHKERRQ(ierr); 10216214f412SHong Zhang ierr = MatSetSizes(mat_elemental,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 10226214f412SHong Zhang ierr = MatSetType(mat_elemental,MATELEMENTAL);CHKERRQ(ierr); 10236214f412SHong Zhang ierr = MatSetUp(mat_elemental);CHKERRQ(ierr); 10246214f412SHong Zhang ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr); 10256214f412SHong Zhang for (row=rstart; row<rend; row++) { 10266214f412SHong Zhang ierr = MatGetRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 10276214f412SHong Zhang /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */ 10286214f412SHong Zhang ierr = MatSetValues(mat_elemental,1,&row,ncols,cols,vals,ADD_VALUES);CHKERRQ(ierr); 10296214f412SHong Zhang for (j=0; j<ncols; j++) { /* lower triangular part */ 10306214f412SHong Zhang if (cols[j] == row) continue; 10316214f412SHong Zhang ierr = MatSetValues(mat_elemental,1,&cols[j],1,&row,&vals[j],ADD_VALUES);CHKERRQ(ierr); 10326214f412SHong Zhang } 10336214f412SHong Zhang ierr = MatRestoreRow(A,row,&ncols,&cols,&vals);CHKERRQ(ierr); 10346214f412SHong Zhang } 10356214f412SHong Zhang ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr); 10366214f412SHong Zhang ierr = MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10376214f412SHong Zhang ierr = MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10386214f412SHong Zhang 1039511c6705SHong Zhang if (reuse == MAT_INPLACE_MATRIX) { 104028be2f97SBarry Smith ierr = MatHeaderReplace(A,&mat_elemental);CHKERRQ(ierr); 10416214f412SHong Zhang } else { 10426214f412SHong Zhang *newmat = mat_elemental; 10436214f412SHong Zhang } 10446214f412SHong Zhang PetscFunctionReturn(0); 10456214f412SHong Zhang } 10466214f412SHong Zhang 1047db31f6deSJed Brown static PetscErrorCode MatDestroy_Elemental(Mat A) 1048db31f6deSJed Brown { 1049db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 1050db31f6deSJed Brown PetscErrorCode ierr; 10515e9f5b67SHong Zhang Mat_Elemental_Grid *commgrid; 10525e9f5b67SHong Zhang PetscBool flg; 10535e9f5b67SHong Zhang MPI_Comm icomm; 1054db31f6deSJed Brown 1055db31f6deSJed Brown PetscFunctionBegin; 1056db31f6deSJed Brown delete a->emat; 1057ea394475SSatish Balay delete a->P; 1058ea394475SSatish Balay delete a->Q; 10595e9f5b67SHong Zhang 1060e8146892SSatish Balay El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A)); 10610c18141cSBarry Smith ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr); 10625e9f5b67SHong Zhang ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr); 10635e9f5b67SHong Zhang if (--commgrid->grid_refct == 0) { 10645e9f5b67SHong Zhang delete commgrid->grid; 10655e9f5b67SHong Zhang ierr = PetscFree(commgrid);CHKERRQ(ierr); 1066a000e53bSHong Zhang ierr = MPI_Keyval_free(&Petsc_Elemental_keyval);CHKERRQ(ierr); 10675e9f5b67SHong Zhang } 10685e9f5b67SHong Zhang ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr); 1069bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",NULL);CHKERRQ(ierr); 1070bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);CHKERRQ(ierr); 1071db31f6deSJed Brown ierr = PetscFree(A->data);CHKERRQ(ierr); 1072db31f6deSJed Brown PetscFunctionReturn(0); 1073db31f6deSJed Brown } 1074db31f6deSJed Brown 1075db31f6deSJed Brown PetscErrorCode MatSetUp_Elemental(Mat A) 1076db31f6deSJed Brown { 1077db31f6deSJed Brown Mat_Elemental *a = (Mat_Elemental*)A->data; 1078db31f6deSJed Brown PetscErrorCode ierr; 1079db31f6deSJed Brown PetscMPIInt rsize,csize; 1080db31f6deSJed Brown 1081db31f6deSJed Brown PetscFunctionBegin; 1082db31f6deSJed Brown ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr); 1083db31f6deSJed Brown ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr); 1084db31f6deSJed Brown 1085efb79153SJack Poulson a->emat->Resize(A->rmap->N,A->cmap->N);CHKERRQ(ierr); 1086e8146892SSatish Balay El::Zero(*a->emat); 1087db31f6deSJed Brown 1088db31f6deSJed Brown ierr = MPI_Comm_size(A->rmap->comm,&rsize);CHKERRQ(ierr); 1089db31f6deSJed Brown ierr = MPI_Comm_size(A->cmap->comm,&csize);CHKERRQ(ierr); 1090ce94432eSBarry Smith if (csize != rsize) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot use row and column communicators of different sizes"); 1091db31f6deSJed Brown a->commsize = rsize; 1092db31f6deSJed Brown a->mr[0] = A->rmap->N % rsize; if (!a->mr[0]) a->mr[0] = rsize; 1093db31f6deSJed Brown a->mr[1] = A->cmap->N % csize; if (!a->mr[1]) a->mr[1] = csize; 1094db31f6deSJed Brown a->m[0] = A->rmap->N / rsize + (a->mr[0] != rsize); 1095db31f6deSJed Brown a->m[1] = A->cmap->N / csize + (a->mr[1] != csize); 1096db31f6deSJed Brown PetscFunctionReturn(0); 1097db31f6deSJed Brown } 1098db31f6deSJed Brown 1099aae2c449SHong Zhang PetscErrorCode MatAssemblyBegin_Elemental(Mat A, MatAssemblyType type) 1100aae2c449SHong Zhang { 1101aae2c449SHong Zhang Mat_Elemental *a = (Mat_Elemental*)A->data; 1102aae2c449SHong Zhang 1103aae2c449SHong Zhang PetscFunctionBegin; 1104fbbcb730SJack Poulson /* printf("Calling ProcessQueues\n"); */ 1105fbbcb730SJack Poulson a->emat->ProcessQueues(); 1106fbbcb730SJack Poulson /* printf("Finished ProcessQueues\n"); */ 1107aae2c449SHong Zhang PetscFunctionReturn(0); 1108aae2c449SHong Zhang } 1109aae2c449SHong Zhang 1110aae2c449SHong Zhang PetscErrorCode MatAssemblyEnd_Elemental(Mat A, MatAssemblyType type) 1111aae2c449SHong Zhang { 1112aae2c449SHong Zhang PetscFunctionBegin; 1113aae2c449SHong Zhang /* Currently does nothing */ 1114aae2c449SHong Zhang PetscFunctionReturn(0); 1115aae2c449SHong Zhang } 1116aae2c449SHong Zhang 11177b30e0f1SHong Zhang PetscErrorCode MatLoad_Elemental(Mat newMat, PetscViewer viewer) 11187b30e0f1SHong Zhang { 11197b30e0f1SHong Zhang PetscErrorCode ierr; 11207b30e0f1SHong Zhang Mat Adense,Ae; 11217b30e0f1SHong Zhang MPI_Comm comm; 11227b30e0f1SHong Zhang 11237b30e0f1SHong Zhang PetscFunctionBegin; 11247b30e0f1SHong Zhang ierr = PetscObjectGetComm((PetscObject)newMat,&comm);CHKERRQ(ierr); 11257b30e0f1SHong Zhang ierr = MatCreate(comm,&Adense);CHKERRQ(ierr); 11267b30e0f1SHong Zhang ierr = MatSetType(Adense,MATDENSE);CHKERRQ(ierr); 11277b30e0f1SHong Zhang ierr = MatLoad(Adense,viewer);CHKERRQ(ierr); 11287b30e0f1SHong Zhang ierr = MatConvert(Adense, MATELEMENTAL, MAT_INITIAL_MATRIX,&Ae);CHKERRQ(ierr); 11297b30e0f1SHong Zhang ierr = MatDestroy(&Adense);CHKERRQ(ierr); 113028be2f97SBarry Smith ierr = MatHeaderReplace(newMat,&Ae);CHKERRQ(ierr); 11317b30e0f1SHong Zhang PetscFunctionReturn(0); 11327b30e0f1SHong Zhang } 11337b30e0f1SHong Zhang 113440d92e34SHong Zhang /* -------------------------------------------------------------------*/ 113540d92e34SHong Zhang static struct _MatOps MatOps_Values = { 113640d92e34SHong Zhang MatSetValues_Elemental, 113740d92e34SHong Zhang 0, 113840d92e34SHong Zhang 0, 113940d92e34SHong Zhang MatMult_Elemental, 114040d92e34SHong Zhang /* 4*/ MatMultAdd_Elemental, 11419426833fSXuan Zhou MatMultTranspose_Elemental, 1142e883f9d5SXuan Zhou MatMultTransposeAdd_Elemental, 114340d92e34SHong Zhang MatSolve_Elemental, 1144df311e6cSXuan Zhou MatSolveAdd_Elemental, 11452ef15aa2SHong Zhang 0, 11462ef15aa2SHong Zhang /*10*/ 0, 114740d92e34SHong Zhang MatLUFactor_Elemental, 114840d92e34SHong Zhang MatCholeskyFactor_Elemental, 114940d92e34SHong Zhang 0, 115040d92e34SHong Zhang MatTranspose_Elemental, 115140d92e34SHong Zhang /*15*/ MatGetInfo_Elemental, 115240d92e34SHong Zhang 0, 115361119200SXuan Zhou MatGetDiagonal_Elemental, 1154ade3cc5eSXuan Zhou MatDiagonalScale_Elemental, 115540d92e34SHong Zhang MatNorm_Elemental, 115640d92e34SHong Zhang /*20*/ MatAssemblyBegin_Elemental, 115740d92e34SHong Zhang MatAssemblyEnd_Elemental, 115832d7a744SHong Zhang MatSetOption_Elemental, 115940d92e34SHong Zhang MatZeroEntries_Elemental, 116040d92e34SHong Zhang /*24*/ 0, 116140d92e34SHong Zhang MatLUFactorSymbolic_Elemental, 116240d92e34SHong Zhang MatLUFactorNumeric_Elemental, 116340d92e34SHong Zhang MatCholeskyFactorSymbolic_Elemental, 116440d92e34SHong Zhang MatCholeskyFactorNumeric_Elemental, 116540d92e34SHong Zhang /*29*/ MatSetUp_Elemental, 116640d92e34SHong Zhang 0, 116740d92e34SHong Zhang 0, 116840d92e34SHong Zhang 0, 116940d92e34SHong Zhang 0, 1170df311e6cSXuan Zhou /*34*/ MatDuplicate_Elemental, 117140d92e34SHong Zhang 0, 117240d92e34SHong Zhang 0, 117340d92e34SHong Zhang 0, 117440d92e34SHong Zhang 0, 117540d92e34SHong Zhang /*39*/ MatAXPY_Elemental, 117640d92e34SHong Zhang 0, 117740d92e34SHong Zhang 0, 117840d92e34SHong Zhang 0, 117940d92e34SHong Zhang MatCopy_Elemental, 118040d92e34SHong Zhang /*44*/ 0, 118140d92e34SHong Zhang MatScale_Elemental, 11827d68702bSBarry Smith MatShift_Basic, 118340d92e34SHong Zhang 0, 118440d92e34SHong Zhang 0, 118540d92e34SHong Zhang /*49*/ 0, 118640d92e34SHong Zhang 0, 118740d92e34SHong Zhang 0, 118840d92e34SHong Zhang 0, 118940d92e34SHong Zhang 0, 119040d92e34SHong Zhang /*54*/ 0, 119140d92e34SHong Zhang 0, 119240d92e34SHong Zhang 0, 119340d92e34SHong Zhang 0, 119440d92e34SHong Zhang 0, 119540d92e34SHong Zhang /*59*/ 0, 119640d92e34SHong Zhang MatDestroy_Elemental, 119740d92e34SHong Zhang MatView_Elemental, 119840d92e34SHong Zhang 0, 119940d92e34SHong Zhang 0, 120040d92e34SHong Zhang /*64*/ 0, 120140d92e34SHong Zhang 0, 120240d92e34SHong Zhang 0, 120340d92e34SHong Zhang 0, 120440d92e34SHong Zhang 0, 120540d92e34SHong Zhang /*69*/ 0, 120640d92e34SHong Zhang 0, 12072ef0cf24SXuan Zhou MatConvert_Elemental_Dense, 120840d92e34SHong Zhang 0, 120940d92e34SHong Zhang 0, 121040d92e34SHong Zhang /*74*/ 0, 121140d92e34SHong Zhang 0, 121240d92e34SHong Zhang 0, 121340d92e34SHong Zhang 0, 121440d92e34SHong Zhang 0, 121540d92e34SHong Zhang /*79*/ 0, 121640d92e34SHong Zhang 0, 121740d92e34SHong Zhang 0, 121840d92e34SHong Zhang 0, 12197b30e0f1SHong Zhang MatLoad_Elemental, 122040d92e34SHong Zhang /*84*/ 0, 122140d92e34SHong Zhang 0, 122240d92e34SHong Zhang 0, 122340d92e34SHong Zhang 0, 122440d92e34SHong Zhang 0, 122540d92e34SHong Zhang /*89*/ MatMatMult_Elemental, 122640d92e34SHong Zhang MatMatMultSymbolic_Elemental, 122740d92e34SHong Zhang MatMatMultNumeric_Elemental, 122840d92e34SHong Zhang 0, 122940d92e34SHong Zhang 0, 123040d92e34SHong Zhang /*94*/ 0, 1231df311e6cSXuan Zhou MatMatTransposeMult_Elemental, 1232df311e6cSXuan Zhou MatMatTransposeMultSymbolic_Elemental, 1233df311e6cSXuan Zhou MatMatTransposeMultNumeric_Elemental, 123440d92e34SHong Zhang 0, 123540d92e34SHong Zhang /*99*/ 0, 123640d92e34SHong Zhang 0, 123740d92e34SHong Zhang 0, 1238dfcb0403SXuan Zhou MatConjugate_Elemental, 123940d92e34SHong Zhang 0, 124040d92e34SHong Zhang /*104*/0, 124140d92e34SHong Zhang 0, 124240d92e34SHong Zhang 0, 124340d92e34SHong Zhang 0, 124440d92e34SHong Zhang 0, 124540d92e34SHong Zhang /*109*/MatMatSolve_Elemental, 124640d92e34SHong Zhang 0, 124740d92e34SHong Zhang 0, 124840d92e34SHong Zhang 0, 124940d92e34SHong Zhang 0, 125040d92e34SHong Zhang /*114*/0, 125140d92e34SHong Zhang 0, 125240d92e34SHong Zhang 0, 125340d92e34SHong Zhang 0, 125440d92e34SHong Zhang 0, 125540d92e34SHong Zhang /*119*/0, 12564a29722dSXuan Zhou MatHermitianTranspose_Elemental, 125740d92e34SHong Zhang 0, 125840d92e34SHong Zhang 0, 125940d92e34SHong Zhang 0, 126040d92e34SHong Zhang /*124*/0, 126140d92e34SHong Zhang 0, 126240d92e34SHong Zhang 0, 126340d92e34SHong Zhang 0, 126440d92e34SHong Zhang 0, 126540d92e34SHong Zhang /*129*/0, 126640d92e34SHong Zhang 0, 126740d92e34SHong Zhang 0, 126840d92e34SHong Zhang 0, 126940d92e34SHong Zhang 0, 127040d92e34SHong Zhang /*134*/0, 127140d92e34SHong Zhang 0, 127240d92e34SHong Zhang 0, 127340d92e34SHong Zhang 0, 127440d92e34SHong Zhang 0 127540d92e34SHong Zhang }; 127640d92e34SHong Zhang 1277ed36708cSHong Zhang /*MC 1278ed36708cSHong Zhang MATELEMENTAL = "elemental" - A matrix type for dense matrices using the Elemental package 1279ed36708cSHong Zhang 1280c2b89b5dSBarry Smith Use ./configure --download-elemental to install PETSc to use Elemental 1281c2b89b5dSBarry Smith 1282c2b89b5dSBarry Smith Use -pc_type lu -pc_factor_mat_solver_package elemental to us this direct solver 1283c2b89b5dSBarry Smith 1284ed36708cSHong Zhang Options Database Keys: 12855cc86fc1SJed Brown + -mat_type elemental - sets the matrix type to "elemental" during a call to MatSetFromOptions() 12865cc86fc1SJed Brown - -mat_elemental_grid_height - sets Grid Height for 2D cyclic ordering of internal matrix 1287ed36708cSHong Zhang 1288ed36708cSHong Zhang Level: beginner 1289ed36708cSHong Zhang 12905cb544a0SHong Zhang .seealso: MATDENSE 1291ed36708cSHong Zhang M*/ 12924a29722dSXuan Zhou 12938cc058d9SJed Brown PETSC_EXTERN PetscErrorCode MatCreate_Elemental(Mat A) 1294db31f6deSJed Brown { 1295db31f6deSJed Brown Mat_Elemental *a; 1296db31f6deSJed Brown PetscErrorCode ierr; 12975682a260SJack Poulson PetscBool flg,flg1; 12985e9f5b67SHong Zhang Mat_Elemental_Grid *commgrid; 12995e9f5b67SHong Zhang MPI_Comm icomm; 13005682a260SJack Poulson PetscInt optv1; 1301db31f6deSJed Brown 1302db31f6deSJed Brown PetscFunctionBegin; 1303607a6623SBarry Smith ierr = PetscElementalInitializePackage();CHKERRQ(ierr); 130440d92e34SHong Zhang ierr = PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 130540d92e34SHong Zhang A->insertmode = NOT_SET_VALUES; 1306db31f6deSJed Brown 1307b00a9115SJed Brown ierr = PetscNewLog(A,&a);CHKERRQ(ierr); 1308db31f6deSJed Brown A->data = (void*)a; 1309db31f6deSJed Brown 1310db31f6deSJed Brown /* Set up the elemental matrix */ 1311e8146892SSatish Balay El::mpi::Comm cxxcomm(PetscObjectComm((PetscObject)A)); 13125e9f5b67SHong Zhang 13135e9f5b67SHong Zhang /* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm */ 13145e9f5b67SHong Zhang if (Petsc_Elemental_keyval == MPI_KEYVAL_INVALID) { 13152a808120SBarry Smith ierr = MPI_Keyval_create(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0);CHKERRQ(ierr); 1316a000e53bSHong Zhang /* ierr = MPI_Comm_create_Keyval(MPI_NULL_COPY_FN,MPI_NULL_DELETE_FN,&Petsc_Elemental_keyval,(void*)0); -- new version? */ 13175e9f5b67SHong Zhang } 13180c18141cSBarry Smith ierr = PetscCommDuplicate(cxxcomm.comm,&icomm,NULL);CHKERRQ(ierr); 13195e9f5b67SHong Zhang ierr = MPI_Attr_get(icomm,Petsc_Elemental_keyval,(void**)&commgrid,(int*)&flg);CHKERRQ(ierr); 13205e9f5b67SHong Zhang if (!flg) { 1321b00a9115SJed Brown ierr = PetscNewLog(A,&commgrid);CHKERRQ(ierr); 13225cb544a0SHong Zhang 1323ce94432eSBarry Smith ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"Elemental Options","Mat");CHKERRQ(ierr); 1324e8146892SSatish Balay /* displayed default grid sizes (CommSize,1) are set by us arbitrarily until El::Grid() is called */ 1325e8146892SSatish Balay ierr = PetscOptionsInt("-mat_elemental_grid_height","Grid Height","None",El::mpi::Size(cxxcomm),&optv1,&flg1);CHKERRQ(ierr); 13265682a260SJack Poulson if (flg1) { 1327e8146892SSatish Balay if (El::mpi::Size(cxxcomm) % optv1 != 0) { 1328e8146892SSatish Balay SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Grid Height %D must evenly divide CommSize %D",optv1,(PetscInt)El::mpi::Size(cxxcomm)); 1329ed667823SXuan Zhou } 1330e8146892SSatish Balay commgrid->grid = new El::Grid(cxxcomm,optv1); /* use user-provided grid height */ 13312ef0cf24SXuan Zhou } else { 1332e8146892SSatish Balay commgrid->grid = new El::Grid(cxxcomm); /* use Elemental default grid sizes */ 1333da0640a4SHong Zhang /* printf("new commgrid->grid = %p\n",commgrid->grid); -- memory leak revealed by valgrind? */ 1334ed667823SXuan Zhou } 13355e9f5b67SHong Zhang commgrid->grid_refct = 1; 13365e9f5b67SHong Zhang ierr = MPI_Attr_put(icomm,Petsc_Elemental_keyval,(void*)commgrid);CHKERRQ(ierr); 1337ea394475SSatish Balay 1338ea394475SSatish Balay a->pivoting = 1; 1339ea394475SSatish Balay ierr = PetscOptionsInt("-mat_elemental_pivoting","Pivoting","None",a->pivoting,&a->pivoting,NULL);CHKERRQ(ierr); 1340ea394475SSatish Balay 13412a808120SBarry Smith ierr = PetscOptionsEnd();CHKERRQ(ierr); 13425e9f5b67SHong Zhang } else { 13435e9f5b67SHong Zhang commgrid->grid_refct++; 13445e9f5b67SHong Zhang } 13455e9f5b67SHong Zhang ierr = PetscCommDestroy(&icomm);CHKERRQ(ierr); 13465e9f5b67SHong Zhang a->grid = commgrid->grid; 1347e8146892SSatish Balay a->emat = new El::DistMatrix<PetscElemScalar>(*a->grid); 134832d7a744SHong Zhang a->roworiented = PETSC_TRUE; 1349db31f6deSJed Brown 1350bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)A,"MatGetOwnershipIS_C",MatGetOwnershipIS_Elemental);CHKERRQ(ierr); 1351db31f6deSJed Brown ierr = PetscObjectChangeTypeName((PetscObject)A,MATELEMENTAL);CHKERRQ(ierr); 1352db31f6deSJed Brown PetscFunctionReturn(0); 1353db31f6deSJed Brown } 1354