1dba47a55SKris Buschelman 24b9ad928SBarry Smith /* 33f457be1SHong Zhang This file defines a "solve the problem redundantly on each subgroup of processor" preconditioner. 44b9ad928SBarry Smith */ 5af0996ceSBarry Smith #include <petsc/private/pcimpl.h> 6c6db04a5SJed Brown #include <petscksp.h> /*I "petscksp.h" I*/ 74b9ad928SBarry Smith 84b9ad928SBarry Smith typedef struct { 93e065800SHong Zhang KSP ksp; 104b9ad928SBarry Smith PC pc; /* actual preconditioner used on each processor */ 11ce94432eSBarry Smith Vec xsub,ysub; /* vectors of a subcommunicator to hold parallel vectors of PetscObjectComm((PetscObject)pc) */ 123f457be1SHong Zhang Vec xdup,ydup; /* parallel vector that congregates xsub or ysub facilitating vector scattering */ 13b3804887SHong Zhang Mat pmats; /* matrix and optional preconditioner matrix belong to a subcommunicator */ 143f457be1SHong Zhang VecScatter scatterin,scatterout; /* scatter used to move all values to each processor group (subcommunicator) */ 15ace3abfcSBarry Smith PetscBool useparallelmat; 16c540e29cSHong Zhang PetscSubcomm psubcomm; 171fbd8f88SHong Zhang PetscInt nsubcomm; /* num of data structure PetscSubcomm */ 18753b7fb9SBarry Smith PetscBool shifttypeset; 19753b7fb9SBarry Smith MatFactorShiftType shifttype; 204b9ad928SBarry Smith } PC_Redundant; 214b9ad928SBarry Smith 22753b7fb9SBarry Smith PetscErrorCode PCFactorSetShiftType_Redundant(PC pc,MatFactorShiftType shifttype) 23753b7fb9SBarry Smith { 24753b7fb9SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 25753b7fb9SBarry Smith PetscErrorCode ierr; 26753b7fb9SBarry Smith 27753b7fb9SBarry Smith PetscFunctionBegin; 28753b7fb9SBarry Smith if (red->ksp) { 29753b7fb9SBarry Smith PC pc; 30753b7fb9SBarry Smith ierr = KSPGetPC(red->ksp,&pc);CHKERRQ(ierr); 31753b7fb9SBarry Smith ierr = PCFactorSetShiftType(pc,shifttype);CHKERRQ(ierr); 32753b7fb9SBarry Smith } else { 33753b7fb9SBarry Smith red->shifttypeset = PETSC_TRUE; 34753b7fb9SBarry Smith red->shifttype = shifttype; 35753b7fb9SBarry Smith } 36753b7fb9SBarry Smith PetscFunctionReturn(0); 37753b7fb9SBarry Smith } 38753b7fb9SBarry Smith 396849ba73SBarry Smith static PetscErrorCode PCView_Redundant(PC pc,PetscViewer viewer) 404b9ad928SBarry Smith { 414b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 42dfbe8321SBarry Smith PetscErrorCode ierr; 43ace3abfcSBarry Smith PetscBool iascii,isstring; 4403ccd0b4SBarry Smith PetscViewer subviewer; 454b9ad928SBarry Smith 464b9ad928SBarry Smith PetscFunctionBegin; 47251f4c67SDmitry Karpeev ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 48251f4c67SDmitry Karpeev ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSTRING,&isstring);CHKERRQ(ierr); 4932077d6dSBarry Smith if (iascii) { 5003ccd0b4SBarry Smith if (!red->psubcomm) { 51efd4aadfSBarry Smith ierr = PetscViewerASCIIPrintf(viewer," Not yet setup\n");CHKERRQ(ierr); 5203ccd0b4SBarry Smith } else { 53efd4aadfSBarry Smith ierr = PetscViewerASCIIPrintf(viewer," First (color=0) of %D PCs follows\n",red->nsubcomm);CHKERRQ(ierr); 543f08860eSBarry Smith ierr = PetscViewerGetSubViewer(viewer,((PetscObject)red->pc)->comm,&subviewer);CHKERRQ(ierr); 55f5dd71faSBarry Smith if (!red->psubcomm->color) { /* only view first redundant pc */ 561575c14dSBarry Smith ierr = PetscViewerASCIIPushTab(subviewer);CHKERRQ(ierr); 573e065800SHong Zhang ierr = KSPView(red->ksp,subviewer);CHKERRQ(ierr); 581575c14dSBarry Smith ierr = PetscViewerASCIIPopTab(subviewer);CHKERRQ(ierr); 594b9ad928SBarry Smith } 603f08860eSBarry Smith ierr = PetscViewerRestoreSubViewer(viewer,((PetscObject)red->pc)->comm,&subviewer);CHKERRQ(ierr); 614b9ad928SBarry Smith } 6203ccd0b4SBarry Smith } else if (isstring) { 6303ccd0b4SBarry Smith ierr = PetscViewerStringSPrintf(viewer," Redundant solver preconditioner");CHKERRQ(ierr); 644b9ad928SBarry Smith } 654b9ad928SBarry Smith PetscFunctionReturn(0); 664b9ad928SBarry Smith } 674b9ad928SBarry Smith 68*19b3b6edSHong Zhang #include <../src/mat/impls/aij/mpi/mpiaij.h> 696849ba73SBarry Smith static PetscErrorCode PCSetUp_Redundant(PC pc) 704b9ad928SBarry Smith { 714b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 72dfbe8321SBarry Smith PetscErrorCode ierr; 731b81debcSHong Zhang PetscInt mstart,mend,mlocal,M; 7413f74950SBarry Smith PetscMPIInt size; 75ce94432eSBarry Smith MPI_Comm comm,subcomm; 76ddc54837SHong Zhang Vec x; 773f457be1SHong Zhang 784b9ad928SBarry Smith PetscFunctionBegin; 79ce94432eSBarry Smith ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 80ddc54837SHong Zhang 81ddc54837SHong Zhang /* if pmatrix set by user is sequential then we do not need to gather the parallel matrix */ 82ddc54837SHong Zhang ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 83ddc54837SHong Zhang if (size == 1) red->useparallelmat = PETSC_FALSE; 841fbd8f88SHong Zhang 854b9ad928SBarry Smith if (!pc->setupcalled) { 861b81debcSHong Zhang PetscInt mloc_sub; 8775024027SHong Zhang if (!red->psubcomm) { /* create red->psubcomm, new ksp and pc over subcomm */ 8875024027SHong Zhang KSP ksp; 8975024027SHong Zhang ierr = PCRedundantGetKSP(pc,&ksp);CHKERRQ(ierr); 901b81debcSHong Zhang } 9175024027SHong Zhang subcomm = PetscSubcommChild(red->psubcomm); 921fbd8f88SHong Zhang 931b81debcSHong Zhang if (red->useparallelmat) { 941b81debcSHong Zhang /* grab the parallel matrix and put it into processors of a subcomminicator */ 9553cd1579SHong Zhang ierr = MatCreateRedundantMatrix(pc->pmat,red->psubcomm->n,subcomm,MAT_INITIAL_MATRIX,&red->pmats);CHKERRQ(ierr); 96b85f2e9bSHong Zhang 97b85f2e9bSHong Zhang ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 98b85f2e9bSHong Zhang if (size > 1) { 99b85f2e9bSHong Zhang PetscBool foundpack; 100*19b3b6edSHong Zhang 101*19b3b6edSHong Zhang PetscBool ismpiaij; 102*19b3b6edSHong Zhang ierr = PetscObjectTypeCompare((PetscObject)(red->pmats),MATMPIAIJ,&ismpiaij);CHKERRQ(ierr); 103*19b3b6edSHong Zhang if (ismpiaij) { 104*19b3b6edSHong Zhang Mat_MPIAIJ *a=(Mat_MPIAIJ*)(red->pmats)->data; 105*19b3b6edSHong Zhang if (!a->Mvctx_mpi1) { /* create a->Mvctx_mpi1 to be used for MatMult() */ 106*19b3b6edSHong Zhang a->Mvctx_mpi1_flg = PETSC_TRUE; 107*19b3b6edSHong Zhang ierr = MatSetUpMultiply_MPIAIJ(red->pmats);CHKERRQ(ierr); 108*19b3b6edSHong Zhang } 109*19b3b6edSHong Zhang } 110*19b3b6edSHong Zhang 111b85f2e9bSHong Zhang ierr = MatGetFactorAvailable(red->pmats,NULL,MAT_FACTOR_LU,&foundpack);CHKERRQ(ierr); 112b85f2e9bSHong Zhang if (!foundpack) { /* reset default ksp and pc */ 113b85f2e9bSHong Zhang ierr = KSPSetType(red->ksp,KSPGMRES);CHKERRQ(ierr); 114b85f2e9bSHong Zhang ierr = PCSetType(red->pc,PCBJACOBI);CHKERRQ(ierr); 115c1619fb6SBarry Smith } else { 116c1619fb6SBarry Smith ierr = PCFactorSetMatSolverPackage(red->pc,NULL);CHKERRQ(ierr); 117b85f2e9bSHong Zhang } 118b85f2e9bSHong Zhang } 119b85f2e9bSHong Zhang 12023ee1639SBarry Smith ierr = KSPSetOperators(red->ksp,red->pmats,red->pmats);CHKERRQ(ierr); 1214b9ad928SBarry Smith 1221b81debcSHong Zhang /* get working vectors xsub and ysub */ 1232a7a6963SBarry Smith ierr = MatCreateVecs(red->pmats,&red->xsub,&red->ysub);CHKERRQ(ierr); 1242fa5cd67SKarl Rupp 1258b96b0d2SHong Zhang /* create working vectors xdup and ydup. 1268b96b0d2SHong Zhang xdup concatenates all xsub's contigously to form a mpi vector over dupcomm (see PetscSubcommCreate_interlaced()) 1278b96b0d2SHong Zhang ydup concatenates all ysub and has empty local arrays because ysub's arrays will be place into it. 128ce94432eSBarry Smith Note: we use communicator dupcomm, not PetscObjectComm((PetscObject)pc)! */ 1291b81debcSHong Zhang ierr = MatGetLocalSize(red->pmats,&mloc_sub,NULL);CHKERRQ(ierr); 13036be1a5eSBarry Smith ierr = VecCreateMPI(PetscSubcommContiguousParent(red->psubcomm),mloc_sub,PETSC_DECIDE,&red->xdup);CHKERRQ(ierr); 13136be1a5eSBarry Smith ierr = VecCreateMPIWithArray(PetscSubcommContiguousParent(red->psubcomm),1,mloc_sub,PETSC_DECIDE,NULL,&red->ydup);CHKERRQ(ierr); 1323f457be1SHong Zhang 133f68be91cSHong Zhang /* create vecscatters */ 134f68be91cSHong Zhang if (!red->scatterin) { /* efficiency of scatterin is independent from psubcomm_type! */ 1353f457be1SHong Zhang IS is1,is2; 1363f457be1SHong Zhang PetscInt *idx1,*idx2,i,j,k; 13745fc02eaSBarry Smith 1382a7a6963SBarry Smith ierr = MatCreateVecs(pc->pmat,&x,0);CHKERRQ(ierr); 1391b81debcSHong Zhang ierr = VecGetSize(x,&M);CHKERRQ(ierr); 1401b81debcSHong Zhang ierr = VecGetOwnershipRange(x,&mstart,&mend);CHKERRQ(ierr); 1411b81debcSHong Zhang mlocal = mend - mstart; 142dcca6d9dSJed Brown ierr = PetscMalloc2(red->psubcomm->n*mlocal,&idx1,red->psubcomm->n*mlocal,&idx2);CHKERRQ(ierr); 1433f457be1SHong Zhang j = 0; 1441fbd8f88SHong Zhang for (k=0; k<red->psubcomm->n; k++) { 1453f457be1SHong Zhang for (i=mstart; i<mend; i++) { 1463f457be1SHong Zhang idx1[j] = i; 147ddc54837SHong Zhang idx2[j++] = i + M*k; 1483f457be1SHong Zhang } 1493f457be1SHong Zhang } 15070b3c8c7SBarry Smith ierr = ISCreateGeneral(comm,red->psubcomm->n*mlocal,idx1,PETSC_COPY_VALUES,&is1);CHKERRQ(ierr); 15170b3c8c7SBarry Smith ierr = ISCreateGeneral(comm,red->psubcomm->n*mlocal,idx2,PETSC_COPY_VALUES,&is2);CHKERRQ(ierr); 152ddc54837SHong Zhang ierr = VecScatterCreate(x,is1,red->xdup,is2,&red->scatterin);CHKERRQ(ierr); 153fcfd50ebSBarry Smith ierr = ISDestroy(&is1);CHKERRQ(ierr); 154fcfd50ebSBarry Smith ierr = ISDestroy(&is2);CHKERRQ(ierr); 1553f457be1SHong Zhang 1566909748bSHong Zhang /* Impl below is good for PETSC_SUBCOMM_INTERLACED (no inter-process communication) and PETSC_SUBCOMM_CONTIGUOUS (communication within subcomm) */ 157ddc54837SHong Zhang ierr = ISCreateStride(comm,mlocal,mstart+ red->psubcomm->color*M,1,&is1);CHKERRQ(ierr); 1583f457be1SHong Zhang ierr = ISCreateStride(comm,mlocal,mstart,1,&is2);CHKERRQ(ierr); 159ddc54837SHong Zhang ierr = VecScatterCreate(red->xdup,is1,x,is2,&red->scatterout);CHKERRQ(ierr); 160fcfd50ebSBarry Smith ierr = ISDestroy(&is1);CHKERRQ(ierr); 161fcfd50ebSBarry Smith ierr = ISDestroy(&is2);CHKERRQ(ierr); 1621d79065fSBarry Smith ierr = PetscFree2(idx1,idx2);CHKERRQ(ierr); 163ddc54837SHong Zhang ierr = VecDestroy(&x);CHKERRQ(ierr); 1641b81debcSHong Zhang } 165ab661555SHong Zhang } else { /* !red->useparallelmat */ 16623ee1639SBarry Smith ierr = KSPSetOperators(red->ksp,pc->mat,pc->pmat);CHKERRQ(ierr); 1671b81debcSHong Zhang } 168ab661555SHong Zhang } else { /* pc->setupcalled */ 1694b9ad928SBarry Smith if (red->useparallelmat) { 170ab661555SHong Zhang MatReuse reuse; 1711b81debcSHong Zhang /* grab the parallel matrix and put it into processors of a subcomminicator */ 1721b81debcSHong Zhang /*--------------------------------------------------------------------------*/ 173ab661555SHong Zhang if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1744b9ad928SBarry Smith /* destroy old matrices */ 1756bf464f9SBarry Smith ierr = MatDestroy(&red->pmats);CHKERRQ(ierr); 176ab661555SHong Zhang reuse = MAT_INITIAL_MATRIX; 1774b9ad928SBarry Smith } else { 178ab661555SHong Zhang reuse = MAT_REUSE_MATRIX; 179ab661555SHong Zhang } 180306c2d5bSBarry Smith ierr = MatCreateRedundantMatrix(pc->pmat,red->psubcomm->n,PetscSubcommChild(red->psubcomm),reuse,&red->pmats);CHKERRQ(ierr); 18123ee1639SBarry Smith ierr = KSPSetOperators(red->ksp,red->pmats,red->pmats);CHKERRQ(ierr); 182ab661555SHong Zhang } else { /* !red->useparallelmat */ 18323ee1639SBarry Smith ierr = KSPSetOperators(red->ksp,pc->mat,pc->pmat);CHKERRQ(ierr); 1844b9ad928SBarry Smith } 185ab661555SHong Zhang } 1861b81debcSHong Zhang 1870c24e6a1SHong Zhang if (pc->setfromoptionscalled) { 1883e065800SHong Zhang ierr = KSPSetFromOptions(red->ksp);CHKERRQ(ierr); 1890c24e6a1SHong Zhang } 1903e065800SHong Zhang ierr = KSPSetUp(red->ksp);CHKERRQ(ierr); 1914b9ad928SBarry Smith PetscFunctionReturn(0); 1924b9ad928SBarry Smith } 1934b9ad928SBarry Smith 1946849ba73SBarry Smith static PetscErrorCode PCApply_Redundant(PC pc,Vec x,Vec y) 1954b9ad928SBarry Smith { 1964b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 197dfbe8321SBarry Smith PetscErrorCode ierr; 1983f457be1SHong Zhang PetscScalar *array; 1994b9ad928SBarry Smith 2004b9ad928SBarry Smith PetscFunctionBegin; 201ddc54837SHong Zhang if (!red->useparallelmat) { 202ddc54837SHong Zhang ierr = KSPSolve(red->ksp,x,y);CHKERRQ(ierr); 203ddc54837SHong Zhang PetscFunctionReturn(0); 204ddc54837SHong Zhang } 205ddc54837SHong Zhang 2063f457be1SHong Zhang /* scatter x to xdup */ 207ca9f406cSSatish Balay ierr = VecScatterBegin(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208ca9f406cSSatish Balay ierr = VecScatterEnd(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2093f457be1SHong Zhang 2103f457be1SHong Zhang /* place xdup's local array into xsub */ 2113f457be1SHong Zhang ierr = VecGetArray(red->xdup,&array);CHKERRQ(ierr); 2123f457be1SHong Zhang ierr = VecPlaceArray(red->xsub,(const PetscScalar*)array);CHKERRQ(ierr); 2134b9ad928SBarry Smith 2144b9ad928SBarry Smith /* apply preconditioner on each processor */ 21583ab6a24SBarry Smith ierr = KSPSolve(red->ksp,red->xsub,red->ysub);CHKERRQ(ierr); 2163f457be1SHong Zhang ierr = VecResetArray(red->xsub);CHKERRQ(ierr); 2173f457be1SHong Zhang ierr = VecRestoreArray(red->xdup,&array);CHKERRQ(ierr); 2184b9ad928SBarry Smith 2193f457be1SHong Zhang /* place ysub's local array into ydup */ 2203f457be1SHong Zhang ierr = VecGetArray(red->ysub,&array);CHKERRQ(ierr); 2213f457be1SHong Zhang ierr = VecPlaceArray(red->ydup,(const PetscScalar*)array);CHKERRQ(ierr); 2223f457be1SHong Zhang 2233f457be1SHong Zhang /* scatter ydup to y */ 224ca9f406cSSatish Balay ierr = VecScatterBegin(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 225ca9f406cSSatish Balay ierr = VecScatterEnd(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2263f457be1SHong Zhang ierr = VecResetArray(red->ydup);CHKERRQ(ierr); 2273f457be1SHong Zhang ierr = VecRestoreArray(red->ysub,&array);CHKERRQ(ierr); 2284b9ad928SBarry Smith PetscFunctionReturn(0); 2294b9ad928SBarry Smith } 2304b9ad928SBarry Smith 231d88bfacbSStefano Zampini static PetscErrorCode PCApplyTranspose_Redundant(PC pc,Vec x,Vec y) 232d88bfacbSStefano Zampini { 233d88bfacbSStefano Zampini PC_Redundant *red = (PC_Redundant*)pc->data; 234d88bfacbSStefano Zampini PetscErrorCode ierr; 235d88bfacbSStefano Zampini PetscScalar *array; 236d88bfacbSStefano Zampini 237d88bfacbSStefano Zampini PetscFunctionBegin; 238d88bfacbSStefano Zampini if (!red->useparallelmat) { 239d88bfacbSStefano Zampini ierr = KSPSolveTranspose(red->ksp,x,y);CHKERRQ(ierr); 240d88bfacbSStefano Zampini PetscFunctionReturn(0); 241d88bfacbSStefano Zampini } 242d88bfacbSStefano Zampini 243d88bfacbSStefano Zampini /* scatter x to xdup */ 244d88bfacbSStefano Zampini ierr = VecScatterBegin(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 245d88bfacbSStefano Zampini ierr = VecScatterEnd(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246d88bfacbSStefano Zampini 247d88bfacbSStefano Zampini /* place xdup's local array into xsub */ 248d88bfacbSStefano Zampini ierr = VecGetArray(red->xdup,&array);CHKERRQ(ierr); 249d88bfacbSStefano Zampini ierr = VecPlaceArray(red->xsub,(const PetscScalar*)array);CHKERRQ(ierr); 250d88bfacbSStefano Zampini 251d88bfacbSStefano Zampini /* apply preconditioner on each processor */ 252d88bfacbSStefano Zampini ierr = KSPSolveTranspose(red->ksp,red->xsub,red->ysub);CHKERRQ(ierr); 253d88bfacbSStefano Zampini ierr = VecResetArray(red->xsub);CHKERRQ(ierr); 254d88bfacbSStefano Zampini ierr = VecRestoreArray(red->xdup,&array);CHKERRQ(ierr); 255d88bfacbSStefano Zampini 256d88bfacbSStefano Zampini /* place ysub's local array into ydup */ 257d88bfacbSStefano Zampini ierr = VecGetArray(red->ysub,&array);CHKERRQ(ierr); 258d88bfacbSStefano Zampini ierr = VecPlaceArray(red->ydup,(const PetscScalar*)array);CHKERRQ(ierr); 259d88bfacbSStefano Zampini 260d88bfacbSStefano Zampini /* scatter ydup to y */ 261d88bfacbSStefano Zampini ierr = VecScatterBegin(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 262d88bfacbSStefano Zampini ierr = VecScatterEnd(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 263d88bfacbSStefano Zampini ierr = VecResetArray(red->ydup);CHKERRQ(ierr); 264d88bfacbSStefano Zampini ierr = VecRestoreArray(red->ysub,&array);CHKERRQ(ierr); 265d88bfacbSStefano Zampini PetscFunctionReturn(0); 266d88bfacbSStefano Zampini } 267d88bfacbSStefano Zampini 2681ea5a559SBarry Smith static PetscErrorCode PCReset_Redundant(PC pc) 2694b9ad928SBarry Smith { 2704b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 271dfbe8321SBarry Smith PetscErrorCode ierr; 2724b9ad928SBarry Smith 2734b9ad928SBarry Smith PetscFunctionBegin; 2741b81debcSHong Zhang if (red->useparallelmat) { 2756bf464f9SBarry Smith ierr = VecScatterDestroy(&red->scatterin);CHKERRQ(ierr); 2766bf464f9SBarry Smith ierr = VecScatterDestroy(&red->scatterout);CHKERRQ(ierr); 2776bf464f9SBarry Smith ierr = VecDestroy(&red->ysub);CHKERRQ(ierr); 2786bf464f9SBarry Smith ierr = VecDestroy(&red->xsub);CHKERRQ(ierr); 2796bf464f9SBarry Smith ierr = VecDestroy(&red->xdup);CHKERRQ(ierr); 2806bf464f9SBarry Smith ierr = VecDestroy(&red->ydup);CHKERRQ(ierr); 2811b81debcSHong Zhang } 2826bf464f9SBarry Smith ierr = MatDestroy(&red->pmats);CHKERRQ(ierr); 2831b81debcSHong Zhang ierr = KSPReset(red->ksp);CHKERRQ(ierr); 2841ea5a559SBarry Smith PetscFunctionReturn(0); 2851ea5a559SBarry Smith } 2861ea5a559SBarry Smith 2871ea5a559SBarry Smith static PetscErrorCode PCDestroy_Redundant(PC pc) 2881ea5a559SBarry Smith { 2891ea5a559SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 2901ea5a559SBarry Smith PetscErrorCode ierr; 2911ea5a559SBarry Smith 2921ea5a559SBarry Smith PetscFunctionBegin; 2931ea5a559SBarry Smith ierr = PCReset_Redundant(pc);CHKERRQ(ierr); 2946bf464f9SBarry Smith ierr = KSPDestroy(&red->ksp);CHKERRQ(ierr); 2956bf464f9SBarry Smith ierr = PetscSubcommDestroy(&red->psubcomm);CHKERRQ(ierr); 296c31cb41cSBarry Smith ierr = PetscFree(pc->data);CHKERRQ(ierr); 2974b9ad928SBarry Smith PetscFunctionReturn(0); 2984b9ad928SBarry Smith } 2994b9ad928SBarry Smith 3004416b707SBarry Smith static PetscErrorCode PCSetFromOptions_Redundant(PetscOptionItems *PetscOptionsObject,PC pc) 3014b9ad928SBarry Smith { 302a98ce0f4SHong Zhang PetscErrorCode ierr; 303a98ce0f4SHong Zhang PC_Redundant *red = (PC_Redundant*)pc->data; 304a98ce0f4SHong Zhang 3054b9ad928SBarry Smith PetscFunctionBegin; 306e55864a3SBarry Smith ierr = PetscOptionsHead(PetscOptionsObject,"Redundant options");CHKERRQ(ierr); 30709a6bc64SHong Zhang ierr = PetscOptionsInt("-pc_redundant_number","Number of redundant pc","PCRedundantSetNumber",red->nsubcomm,&red->nsubcomm,0);CHKERRQ(ierr); 308a98ce0f4SHong Zhang ierr = PetscOptionsTail();CHKERRQ(ierr); 3094b9ad928SBarry Smith PetscFunctionReturn(0); 3104b9ad928SBarry Smith } 3114b9ad928SBarry Smith 312f7a08781SBarry Smith static PetscErrorCode PCRedundantSetNumber_Redundant(PC pc,PetscInt nreds) 31309a6bc64SHong Zhang { 31409a6bc64SHong Zhang PC_Redundant *red = (PC_Redundant*)pc->data; 31509a6bc64SHong Zhang 31609a6bc64SHong Zhang PetscFunctionBegin; 31709a6bc64SHong Zhang red->nsubcomm = nreds; 31809a6bc64SHong Zhang PetscFunctionReturn(0); 31909a6bc64SHong Zhang } 32009a6bc64SHong Zhang 32109a6bc64SHong Zhang /*@ 32209a6bc64SHong Zhang PCRedundantSetNumber - Sets the number of redundant preconditioner contexts. 32309a6bc64SHong Zhang 3243f9fe445SBarry Smith Logically Collective on PC 32509a6bc64SHong Zhang 32609a6bc64SHong Zhang Input Parameters: 32709a6bc64SHong Zhang + pc - the preconditioner context 3289b21d695SBarry Smith - nredundant - number of redundant preconditioner contexts; for example if you are using 64 MPI processes and 3299b21d695SBarry Smith use an nredundant of 4 there will be 4 parallel solves each on 16 = 64/4 processes. 33009a6bc64SHong Zhang 33109a6bc64SHong Zhang Level: advanced 33209a6bc64SHong Zhang 33309a6bc64SHong Zhang .keywords: PC, redundant solve 33409a6bc64SHong Zhang @*/ 3357087cfbeSBarry Smith PetscErrorCode PCRedundantSetNumber(PC pc,PetscInt nredundant) 33609a6bc64SHong Zhang { 3374ac538c5SBarry Smith PetscErrorCode ierr; 33809a6bc64SHong Zhang 33909a6bc64SHong Zhang PetscFunctionBegin; 3400700a824SBarry Smith PetscValidHeaderSpecific(pc,PC_CLASSID,1); 341ce94432eSBarry Smith if (nredundant <= 0) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_WRONG, "num of redundant pc %D must be positive",nredundant); 3424ac538c5SBarry Smith ierr = PetscTryMethod(pc,"PCRedundantSetNumber_C",(PC,PetscInt),(pc,nredundant));CHKERRQ(ierr); 34309a6bc64SHong Zhang PetscFunctionReturn(0); 34409a6bc64SHong Zhang } 34509a6bc64SHong Zhang 346f7a08781SBarry Smith static PetscErrorCode PCRedundantSetScatter_Redundant(PC pc,VecScatter in,VecScatter out) 3474b9ad928SBarry Smith { 3484b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 349dfbe8321SBarry Smith PetscErrorCode ierr; 3504b9ad928SBarry Smith 3514b9ad928SBarry Smith PetscFunctionBegin; 3524b9ad928SBarry Smith ierr = PetscObjectReference((PetscObject)in);CHKERRQ(ierr); 3536bf464f9SBarry Smith ierr = VecScatterDestroy(&red->scatterin);CHKERRQ(ierr); 3542fa5cd67SKarl Rupp 355c3122656SLisandro Dalcin red->scatterin = in; 3562fa5cd67SKarl Rupp 3574b9ad928SBarry Smith ierr = PetscObjectReference((PetscObject)out);CHKERRQ(ierr); 3586bf464f9SBarry Smith ierr = VecScatterDestroy(&red->scatterout);CHKERRQ(ierr); 359c3122656SLisandro Dalcin red->scatterout = out; 3604b9ad928SBarry Smith PetscFunctionReturn(0); 3614b9ad928SBarry Smith } 3624b9ad928SBarry Smith 3634b9ad928SBarry Smith /*@ 3644b9ad928SBarry Smith PCRedundantSetScatter - Sets the scatter used to copy values into the 3654b9ad928SBarry Smith redundant local solve and the scatter to move them back into the global 3664b9ad928SBarry Smith vector. 3674b9ad928SBarry Smith 3683f9fe445SBarry Smith Logically Collective on PC 3694b9ad928SBarry Smith 3704b9ad928SBarry Smith Input Parameters: 3714b9ad928SBarry Smith + pc - the preconditioner context 3724b9ad928SBarry Smith . in - the scatter to move the values in 3734b9ad928SBarry Smith - out - the scatter to move them out 3744b9ad928SBarry Smith 3754b9ad928SBarry Smith Level: advanced 3764b9ad928SBarry Smith 3774b9ad928SBarry Smith .keywords: PC, redundant solve 3784b9ad928SBarry Smith @*/ 3797087cfbeSBarry Smith PetscErrorCode PCRedundantSetScatter(PC pc,VecScatter in,VecScatter out) 3804b9ad928SBarry Smith { 3814ac538c5SBarry Smith PetscErrorCode ierr; 3824b9ad928SBarry Smith 3834b9ad928SBarry Smith PetscFunctionBegin; 3840700a824SBarry Smith PetscValidHeaderSpecific(pc,PC_CLASSID,1); 3850700a824SBarry Smith PetscValidHeaderSpecific(in,VEC_SCATTER_CLASSID,2); 3860700a824SBarry Smith PetscValidHeaderSpecific(out,VEC_SCATTER_CLASSID,3); 3874ac538c5SBarry Smith ierr = PetscTryMethod(pc,"PCRedundantSetScatter_C",(PC,VecScatter,VecScatter),(pc,in,out));CHKERRQ(ierr); 3884b9ad928SBarry Smith PetscFunctionReturn(0); 3894b9ad928SBarry Smith } 3904b9ad928SBarry Smith 391f7a08781SBarry Smith static PetscErrorCode PCRedundantGetKSP_Redundant(PC pc,KSP *innerksp) 3924b9ad928SBarry Smith { 3935f06b7aaSBarry Smith PetscErrorCode ierr; 3944b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 39575024027SHong Zhang MPI_Comm comm,subcomm; 39675024027SHong Zhang const char *prefix; 3974b9ad928SBarry Smith 3984b9ad928SBarry Smith PetscFunctionBegin; 39975024027SHong Zhang if (!red->psubcomm) { 400e5acf8a4SHong Zhang ierr = PCGetOptionsPrefix(pc,&prefix);CHKERRQ(ierr); 401e5acf8a4SHong Zhang 40275024027SHong Zhang ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 40375024027SHong Zhang ierr = PetscSubcommCreate(comm,&red->psubcomm);CHKERRQ(ierr); 40475024027SHong Zhang ierr = PetscSubcommSetNumber(red->psubcomm,red->nsubcomm);CHKERRQ(ierr); 40575024027SHong Zhang ierr = PetscSubcommSetType(red->psubcomm,PETSC_SUBCOMM_CONTIGUOUS);CHKERRQ(ierr); 406e5acf8a4SHong Zhang 407e5acf8a4SHong Zhang ierr = PetscSubcommSetOptionsPrefix(red->psubcomm,prefix);CHKERRQ(ierr); 408e5acf8a4SHong Zhang ierr = PetscSubcommSetFromOptions(red->psubcomm);CHKERRQ(ierr); 40975024027SHong Zhang ierr = PetscLogObjectMemory((PetscObject)pc,sizeof(PetscSubcomm));CHKERRQ(ierr); 41075024027SHong Zhang 41175024027SHong Zhang /* create a new PC that processors in each subcomm have copy of */ 41275024027SHong Zhang subcomm = PetscSubcommChild(red->psubcomm); 41375024027SHong Zhang 41475024027SHong Zhang ierr = KSPCreate(subcomm,&red->ksp);CHKERRQ(ierr); 41575024027SHong Zhang ierr = KSPSetErrorIfNotConverged(red->ksp,pc->erroriffailure);CHKERRQ(ierr); 41675024027SHong Zhang ierr = PetscObjectIncrementTabLevel((PetscObject)red->ksp,(PetscObject)pc,1);CHKERRQ(ierr); 41775024027SHong Zhang ierr = PetscLogObjectParent((PetscObject)pc,(PetscObject)red->ksp);CHKERRQ(ierr); 41875024027SHong Zhang ierr = KSPSetType(red->ksp,KSPPREONLY);CHKERRQ(ierr); 41975024027SHong Zhang ierr = KSPGetPC(red->ksp,&red->pc);CHKERRQ(ierr); 42075024027SHong Zhang ierr = PCSetType(red->pc,PCLU);CHKERRQ(ierr); 421753b7fb9SBarry Smith if (red->shifttypeset) { 422753b7fb9SBarry Smith ierr = PCFactorSetShiftType(red->pc,red->shifttype);CHKERRQ(ierr); 423753b7fb9SBarry Smith red->shifttypeset = PETSC_FALSE; 424753b7fb9SBarry Smith } 42575024027SHong Zhang ierr = KSPSetOptionsPrefix(red->ksp,prefix);CHKERRQ(ierr); 42675024027SHong Zhang ierr = KSPAppendOptionsPrefix(red->ksp,"redundant_");CHKERRQ(ierr); 42775024027SHong Zhang } 42883ab6a24SBarry Smith *innerksp = red->ksp; 4294b9ad928SBarry Smith PetscFunctionReturn(0); 4304b9ad928SBarry Smith } 4314b9ad928SBarry Smith 4324b9ad928SBarry Smith /*@ 43383ab6a24SBarry Smith PCRedundantGetKSP - Gets the less parallel KSP created by the redundant PC. 4344b9ad928SBarry Smith 4354b9ad928SBarry Smith Not Collective 4364b9ad928SBarry Smith 4374b9ad928SBarry Smith Input Parameter: 4384b9ad928SBarry Smith . pc - the preconditioner context 4394b9ad928SBarry Smith 4404b9ad928SBarry Smith Output Parameter: 44183ab6a24SBarry Smith . innerksp - the KSP on the smaller set of processes 4424b9ad928SBarry Smith 4434b9ad928SBarry Smith Level: advanced 4444b9ad928SBarry Smith 4454b9ad928SBarry Smith .keywords: PC, redundant solve 4464b9ad928SBarry Smith @*/ 44783ab6a24SBarry Smith PetscErrorCode PCRedundantGetKSP(PC pc,KSP *innerksp) 4484b9ad928SBarry Smith { 4494ac538c5SBarry Smith PetscErrorCode ierr; 4504b9ad928SBarry Smith 4514b9ad928SBarry Smith PetscFunctionBegin; 4520700a824SBarry Smith PetscValidHeaderSpecific(pc,PC_CLASSID,1); 45383ab6a24SBarry Smith PetscValidPointer(innerksp,2); 454753b7fb9SBarry Smith ierr = PetscUseMethod(pc,"PCRedundantGetKSP_C",(PC,KSP*),(pc,innerksp));CHKERRQ(ierr); 4554b9ad928SBarry Smith PetscFunctionReturn(0); 4564b9ad928SBarry Smith } 4574b9ad928SBarry Smith 458f7a08781SBarry Smith static PetscErrorCode PCRedundantGetOperators_Redundant(PC pc,Mat *mat,Mat *pmat) 4594b9ad928SBarry Smith { 4604b9ad928SBarry Smith PC_Redundant *red = (PC_Redundant*)pc->data; 4614b9ad928SBarry Smith 4624b9ad928SBarry Smith PetscFunctionBegin; 463b3804887SHong Zhang if (mat) *mat = red->pmats; 464b3804887SHong Zhang if (pmat) *pmat = red->pmats; 4654b9ad928SBarry Smith PetscFunctionReturn(0); 4664b9ad928SBarry Smith } 4674b9ad928SBarry Smith 4684b9ad928SBarry Smith /*@ 4694b9ad928SBarry Smith PCRedundantGetOperators - gets the sequential matrix and preconditioner matrix 4704b9ad928SBarry Smith 4714b9ad928SBarry Smith Not Collective 4724b9ad928SBarry Smith 4734b9ad928SBarry Smith Input Parameter: 4744b9ad928SBarry Smith . pc - the preconditioner context 4754b9ad928SBarry Smith 4764b9ad928SBarry Smith Output Parameters: 4774b9ad928SBarry Smith + mat - the matrix 4784b9ad928SBarry Smith - pmat - the (possibly different) preconditioner matrix 4794b9ad928SBarry Smith 4804b9ad928SBarry Smith Level: advanced 4814b9ad928SBarry Smith 4824b9ad928SBarry Smith .keywords: PC, redundant solve 4834b9ad928SBarry Smith @*/ 4847087cfbeSBarry Smith PetscErrorCode PCRedundantGetOperators(PC pc,Mat *mat,Mat *pmat) 4854b9ad928SBarry Smith { 4864ac538c5SBarry Smith PetscErrorCode ierr; 4874b9ad928SBarry Smith 4884b9ad928SBarry Smith PetscFunctionBegin; 4890700a824SBarry Smith PetscValidHeaderSpecific(pc,PC_CLASSID,1); 4904482741eSBarry Smith if (mat) PetscValidPointer(mat,2); 4914482741eSBarry Smith if (pmat) PetscValidPointer(pmat,3); 492163d334eSBarry Smith ierr = PetscUseMethod(pc,"PCRedundantGetOperators_C",(PC,Mat*,Mat*),(pc,mat,pmat));CHKERRQ(ierr); 4934b9ad928SBarry Smith PetscFunctionReturn(0); 4944b9ad928SBarry Smith } 4954b9ad928SBarry Smith 4964b9ad928SBarry Smith /* -------------------------------------------------------------------------------------*/ 49737a17b4dSBarry Smith /*MC 49883ab6a24SBarry Smith PCREDUNDANT - Runs a KSP solver with preconditioner for the entire problem on subgroups of processors 49937a17b4dSBarry Smith 50083ab6a24SBarry Smith Options for the redundant preconditioners can be set with -redundant_pc_xxx for the redundant KSP with -redundant_ksp_xxx 50137a17b4dSBarry Smith 50209391456SBarry Smith Options Database: 5039b21d695SBarry Smith . -pc_redundant_number <n> - number of redundant solves, for example if you are using 64 MPI processes and 5049b21d695SBarry Smith use an n of 4 there will be 4 parallel solves each on 16 = 64/4 processes. 50509391456SBarry Smith 50637a17b4dSBarry Smith Level: intermediate 50737a17b4dSBarry Smith 50883ab6a24SBarry Smith Notes: The default KSP is preonly and the default PC is LU. 50983ab6a24SBarry Smith 510753b7fb9SBarry Smith PCFactorSetShiftType() applied to this PC will convey they shift type into the inner PC if it is factorization based. 511753b7fb9SBarry Smith 51283ab6a24SBarry Smith Developer Notes: Note that PCSetInitialGuessNonzero() is not used by this class but likely should be. 5139cfaa89bSBarry Smith 51437a17b4dSBarry Smith .seealso: PCCreate(), PCSetType(), PCType (for list of available types), PCRedundantSetScatter(), 51583ab6a24SBarry Smith PCRedundantGetKSP(), PCRedundantGetOperators(), PCRedundantSetNumber() 51637a17b4dSBarry Smith M*/ 51737a17b4dSBarry Smith 5188cc058d9SJed Brown PETSC_EXTERN PetscErrorCode PCCreate_Redundant(PC pc) 5194b9ad928SBarry Smith { 520dfbe8321SBarry Smith PetscErrorCode ierr; 5214b9ad928SBarry Smith PC_Redundant *red; 52269db28dcSHong Zhang PetscMPIInt size; 5233f457be1SHong Zhang 5244b9ad928SBarry Smith PetscFunctionBegin; 525b00a9115SJed Brown ierr = PetscNewLog(pc,&red);CHKERRQ(ierr); 526ce94432eSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 5272fa5cd67SKarl Rupp 52869db28dcSHong Zhang red->nsubcomm = size; 5294b9ad928SBarry Smith red->useparallelmat = PETSC_TRUE; 5301fbd8f88SHong Zhang pc->data = (void*)red; 5314b9ad928SBarry Smith 5324b9ad928SBarry Smith pc->ops->apply = PCApply_Redundant; 533d88bfacbSStefano Zampini pc->ops->applytranspose = PCApplyTranspose_Redundant; 5344b9ad928SBarry Smith pc->ops->setup = PCSetUp_Redundant; 5354b9ad928SBarry Smith pc->ops->destroy = PCDestroy_Redundant; 5361ea5a559SBarry Smith pc->ops->reset = PCReset_Redundant; 5374b9ad928SBarry Smith pc->ops->setfromoptions = PCSetFromOptions_Redundant; 5384b9ad928SBarry Smith pc->ops->view = PCView_Redundant; 5392fa5cd67SKarl Rupp 540bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)pc,"PCRedundantSetScatter_C",PCRedundantSetScatter_Redundant);CHKERRQ(ierr); 541bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)pc,"PCRedundantSetNumber_C",PCRedundantSetNumber_Redundant);CHKERRQ(ierr); 542bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)pc,"PCRedundantGetKSP_C",PCRedundantGetKSP_Redundant);CHKERRQ(ierr); 543bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)pc,"PCRedundantGetOperators_C",PCRedundantGetOperators_Redundant);CHKERRQ(ierr); 544753b7fb9SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)pc,"PCFactorSetShiftType_C",PCFactorSetShiftType_Redundant);CHKERRQ(ierr); 5454b9ad928SBarry Smith PetscFunctionReturn(0); 5464b9ad928SBarry Smith } 547b2573a8aSBarry Smith 548