1*dba47a55SKris Buschelman #define PETSCKSP_DLL 26218e575SSatish Balay /* 36218e575SSatish Balay Provides an interface to the Tufo-Fischer parallel direct solver 46218e575SSatish Balay */ 56218e575SSatish Balay 66218e575SSatish Balay #include "src/ksp/pc/pcimpl.h" /*I "petscpc.h" I*/ 76218e575SSatish Balay #include "src/mat/impls/aij/mpi/mpiaij.h" 87758a8cdSBarry Smith #include "src/ksp/pc/impls/tfs/tfs.h" 96218e575SSatish Balay 106218e575SSatish Balay typedef struct { 116218e575SSatish Balay xxt_ADT xxt; 126218e575SSatish Balay xyt_ADT xyt; 136218e575SSatish Balay Vec b,xd,xo; 146218e575SSatish Balay PetscInt nd; 156218e575SSatish Balay } PC_TFS; 166218e575SSatish Balay 176218e575SSatish Balay #undef __FUNCT__ 186218e575SSatish Balay #define __FUNCT__ "PCDestroy_TFS" 196218e575SSatish Balay PetscErrorCode PCDestroy_TFS(PC pc) 206218e575SSatish Balay { 216218e575SSatish Balay PC_TFS *tfs = (PC_TFS*)pc->data; 226218e575SSatish Balay PetscErrorCode ierr; 236218e575SSatish Balay 246218e575SSatish Balay PetscFunctionBegin; 256218e575SSatish Balay /* free the XXT datastructures */ 266218e575SSatish Balay if (tfs->xxt) { 276218e575SSatish Balay ierr = XXT_free(tfs->xxt);CHKERRQ(ierr); 286218e575SSatish Balay } 296218e575SSatish Balay if (tfs->xyt) { 306218e575SSatish Balay ierr = XYT_free(tfs->xyt);CHKERRQ(ierr); 316218e575SSatish Balay } 326218e575SSatish Balay if (tfs->b) { 336218e575SSatish Balay ierr = VecDestroy(tfs->b);CHKERRQ(ierr); 346218e575SSatish Balay } 356218e575SSatish Balay if (tfs->xd) { 366218e575SSatish Balay ierr = VecDestroy(tfs->xd);CHKERRQ(ierr); 376218e575SSatish Balay } 386218e575SSatish Balay if (tfs->xo) { 396218e575SSatish Balay ierr = VecDestroy(tfs->xo);CHKERRQ(ierr); 406218e575SSatish Balay } 416218e575SSatish Balay ierr = PetscFree(tfs);CHKERRQ(ierr); 426218e575SSatish Balay PetscFunctionReturn(0); 436218e575SSatish Balay } 446218e575SSatish Balay 456218e575SSatish Balay #undef __FUNCT__ 466218e575SSatish Balay #define __FUNCT__ "PCApply_TFS_XXT" 476218e575SSatish Balay static PetscErrorCode PCApply_TFS_XXT(PC pc,Vec x,Vec y) 486218e575SSatish Balay { 496218e575SSatish Balay PC_TFS *tfs = (PC_TFS*)pc->data; 506218e575SSatish Balay PetscScalar *xx,*yy; 516218e575SSatish Balay PetscErrorCode ierr; 526218e575SSatish Balay 536218e575SSatish Balay PetscFunctionBegin; 546218e575SSatish Balay ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 556218e575SSatish Balay ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 566218e575SSatish Balay ierr = XXT_solve(tfs->xxt,yy,xx);CHKERRQ(ierr); 576218e575SSatish Balay ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 586218e575SSatish Balay ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 596218e575SSatish Balay PetscFunctionReturn(0); 606218e575SSatish Balay } 616218e575SSatish Balay 626218e575SSatish Balay #undef __FUNCT__ 636218e575SSatish Balay #define __FUNCT__ "PCApply_TFS_XYT" 646218e575SSatish Balay static PetscErrorCode PCApply_TFS_XYT(PC pc,Vec x,Vec y) 656218e575SSatish Balay { 666218e575SSatish Balay PC_TFS *tfs = (PC_TFS*)pc->data; 676218e575SSatish Balay PetscScalar *xx,*yy; 686218e575SSatish Balay PetscErrorCode ierr; 696218e575SSatish Balay 706218e575SSatish Balay PetscFunctionBegin; 716218e575SSatish Balay ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 726218e575SSatish Balay ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 736218e575SSatish Balay ierr = XYT_solve(tfs->xyt,yy,xx);CHKERRQ(ierr); 746218e575SSatish Balay ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 756218e575SSatish Balay ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 766218e575SSatish Balay PetscFunctionReturn(0); 776218e575SSatish Balay } 786218e575SSatish Balay 796218e575SSatish Balay #undef __FUNCT__ 806218e575SSatish Balay #define __FUNCT__ "LocalMult_TFS" 816218e575SSatish Balay static PetscErrorCode LocalMult_TFS(PC pc,PetscScalar *xin,PetscScalar *xout) 826218e575SSatish Balay { 836218e575SSatish Balay PC_TFS *tfs = (PC_TFS*)pc->data; 846218e575SSatish Balay Mat A = pc->pmat; 856218e575SSatish Balay Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 866218e575SSatish Balay PetscErrorCode ierr; 876218e575SSatish Balay 886218e575SSatish Balay PetscFunctionBegin; 896218e575SSatish Balay ierr = VecPlaceArray(tfs->b,xout);CHKERRQ(ierr); 906218e575SSatish Balay ierr = VecPlaceArray(tfs->xd,xin);CHKERRQ(ierr); 916218e575SSatish Balay ierr = VecPlaceArray(tfs->xo,xin+tfs->nd);CHKERRQ(ierr); 926218e575SSatish Balay ierr = MatMult(a->A,tfs->xd,tfs->b);CHKERRQ(ierr); 936218e575SSatish Balay ierr = MatMultAdd(a->B,tfs->xo,tfs->b,tfs->b);CHKERRQ(ierr); 946218e575SSatish Balay PetscFunctionReturn(0); 956218e575SSatish Balay } 966218e575SSatish Balay 976218e575SSatish Balay #undef __FUNCT__ 986218e575SSatish Balay #define __FUNCT__ "PCSetUp_TFS" 996218e575SSatish Balay static PetscErrorCode PCSetUp_TFS(PC pc) 1006218e575SSatish Balay { 1016218e575SSatish Balay PC_TFS *tfs = (PC_TFS*)pc->data; 1026218e575SSatish Balay Mat A = pc->pmat; 1036218e575SSatish Balay Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1046218e575SSatish Balay PetscErrorCode ierr; 1056218e575SSatish Balay PetscInt *localtoglobal,ncol,i; 1066218e575SSatish Balay PetscTruth ismpiaij; 1076218e575SSatish Balay 1086218e575SSatish Balay /* 1096218e575SSatish Balay PetscTruth issymmetric; 1106218e575SSatish Balay Petsc Real tol = 0.0; 1116218e575SSatish Balay */ 1126218e575SSatish Balay 1136218e575SSatish Balay PetscFunctionBegin; 1146218e575SSatish Balay if (A->N != A->M) SETERRQ(PETSC_ERR_ARG_SIZ,"matrix must be square"); 1156218e575SSatish Balay ierr = PetscTypeCompare((PetscObject)pc->pmat,MATMPIAIJ,&ismpiaij);CHKERRQ(ierr); 1166218e575SSatish Balay if (!ismpiaij) { 1176218e575SSatish Balay SETERRQ(PETSC_ERR_SUP,"Currently only supports MPIAIJ matrices"); 1186218e575SSatish Balay } 1196218e575SSatish Balay 1206218e575SSatish Balay /* generate the local to global mapping */ 1216218e575SSatish Balay ncol = a->A->n + a->B->n; 1226218e575SSatish Balay ierr = PetscMalloc((ncol)*sizeof(int),&localtoglobal);CHKERRQ(ierr); 1236218e575SSatish Balay for (i=0; i<a->A->n; i++) { 1246218e575SSatish Balay localtoglobal[i] = a->cstart + i + 1; 1256218e575SSatish Balay } 1266218e575SSatish Balay for (i=0; i<a->B->n; i++) { 1276218e575SSatish Balay localtoglobal[i+a->A->n] = a->garray[i] + 1; 1286218e575SSatish Balay } 1296218e575SSatish Balay /* generate the vectors needed for the local solves */ 1306218e575SSatish Balay ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,a->A->m,PETSC_NULL,&tfs->b);CHKERRQ(ierr); 1316218e575SSatish Balay ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,a->A->n,PETSC_NULL,&tfs->xd);CHKERRQ(ierr); 1326218e575SSatish Balay ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,a->B->n,PETSC_NULL,&tfs->xo);CHKERRQ(ierr); 1336218e575SSatish Balay tfs->nd = a->A->n; 1346218e575SSatish Balay 1356218e575SSatish Balay 1366218e575SSatish Balay /* ierr = MatIsSymmetric(A,tol,&issymmetric); */ 1376218e575SSatish Balay /* if (issymmetric) { */ 138585bbf96SBarry Smith ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 1396218e575SSatish Balay if (A->symmetric) { 1406218e575SSatish Balay tfs->xxt = XXT_new(); 1416218e575SSatish Balay ierr = XXT_factor(tfs->xxt,localtoglobal,A->m,ncol,(void*)LocalMult_TFS,pc);CHKERRQ(ierr); 1426218e575SSatish Balay pc->ops->apply = PCApply_TFS_XXT; 1436218e575SSatish Balay } else { 1446218e575SSatish Balay tfs->xyt = XYT_new(); 1456218e575SSatish Balay ierr = XYT_factor(tfs->xyt,localtoglobal,A->m,ncol,(void*)LocalMult_TFS,pc);CHKERRQ(ierr); 1466218e575SSatish Balay pc->ops->apply = PCApply_TFS_XYT; 1476218e575SSatish Balay } 1486218e575SSatish Balay 1496218e575SSatish Balay ierr = PetscFree(localtoglobal);CHKERRQ(ierr); 1506218e575SSatish Balay PetscFunctionReturn(0); 1516218e575SSatish Balay } 1526218e575SSatish Balay 1536218e575SSatish Balay #undef __FUNCT__ 1546218e575SSatish Balay #define __FUNCT__ "PCSetFromOptions_TFS" 1556218e575SSatish Balay static PetscErrorCode PCSetFromOptions_TFS(PC pc) 1566218e575SSatish Balay { 1576218e575SSatish Balay PetscFunctionBegin; 1586218e575SSatish Balay PetscFunctionReturn(0); 1596218e575SSatish Balay } 1606218e575SSatish Balay #undef __FUNCT__ 1616218e575SSatish Balay #define __FUNCT__ "PCView_TFS" 1626218e575SSatish Balay static PetscErrorCode PCView_TFS(PC pc,PetscViewer viewer) 1636218e575SSatish Balay { 1646218e575SSatish Balay PetscFunctionBegin; 1656218e575SSatish Balay PetscFunctionReturn(0); 1666218e575SSatish Balay } 1676218e575SSatish Balay 1686218e575SSatish Balay EXTERN_C_BEGIN 1696218e575SSatish Balay #undef __FUNCT__ 1706218e575SSatish Balay #define __FUNCT__ "PCCreate_TFS" 171*dba47a55SKris Buschelman PetscErrorCode PETSCKSP_DLLEXPORT PCCreate_TFS(PC pc) 1726218e575SSatish Balay { 1736218e575SSatish Balay PetscErrorCode ierr; 1746218e575SSatish Balay PC_TFS *tfs; 1756218e575SSatish Balay 1766218e575SSatish Balay PetscFunctionBegin; 1776218e575SSatish Balay ierr = PetscNew(PC_TFS,&tfs);CHKERRQ(ierr); 17852e6d16bSBarry Smith ierr = PetscLogObjectMemory(pc,sizeof(PC_TFS));CHKERRQ(ierr); 1796218e575SSatish Balay 1806218e575SSatish Balay tfs->xxt = 0; 1816218e575SSatish Balay tfs->xyt = 0; 1826218e575SSatish Balay tfs->b = 0; 1836218e575SSatish Balay tfs->xd = 0; 1846218e575SSatish Balay tfs->xo = 0; 1856218e575SSatish Balay tfs->nd = 0; 1866218e575SSatish Balay 1876218e575SSatish Balay pc->ops->apply = 0; 1886218e575SSatish Balay pc->ops->applytranspose = 0; 1896218e575SSatish Balay pc->ops->setup = PCSetUp_TFS; 1906218e575SSatish Balay pc->ops->destroy = PCDestroy_TFS; 1916218e575SSatish Balay pc->ops->setfromoptions = PCSetFromOptions_TFS; 1926218e575SSatish Balay pc->ops->view = PCView_TFS; 1936218e575SSatish Balay pc->ops->applyrichardson = 0; 1946218e575SSatish Balay pc->ops->applysymmetricleft = 0; 1956218e575SSatish Balay pc->ops->applysymmetricright = 0; 1966218e575SSatish Balay pc->data = (void*)tfs; 1976218e575SSatish Balay PetscFunctionReturn(0); 1986218e575SSatish Balay } 1996218e575SSatish Balay EXTERN_C_END 2006218e575SSatish Balay 201