134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h> 234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 308122e43SStefano Zampini #include <petscblaslapack.h> 434a97f8cSStefano Zampini 53202ece2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*); 65ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*); 7d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC,Vec,Vec); 8d62866d3SStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolve(PC,Vec,Vec); 9d62866d3SStefano Zampini 10d62866d3SStefano Zampini #undef __FUNCT__ 11e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsCorrectionSolve_Private" 12e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose) 13d62866d3SStefano Zampini { 14d62866d3SStefano Zampini PCBDDCReuseMumps ctx; 15ffbec71fSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 16d62866d3SStefano Zampini PetscInt ival; 17ffbec71fSStefano Zampini #endif 18d62866d3SStefano Zampini PetscErrorCode ierr; 19d62866d3SStefano Zampini 20d62866d3SStefano Zampini PetscFunctionBegin; 21d62866d3SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 22d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 23d62866d3SStefano Zampini ierr = MatMumpsGetIcntl(ctx->F,26,&ival);CHKERRQ(ierr); 24d62866d3SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 25d62866d3SStefano Zampini #endif 26e28d306cSStefano Zampini if (transpose) { 27e28d306cSStefano Zampini ierr = MatSolveTranspose(ctx->F,rhs,sol);CHKERRQ(ierr); 28e28d306cSStefano Zampini } else { 296816873aSStefano Zampini ierr = MatSolve(ctx->F,rhs,sol);CHKERRQ(ierr); 30e28d306cSStefano Zampini } 31d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 32d62866d3SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,ival);CHKERRQ(ierr); 33d62866d3SStefano Zampini #endif 34d62866d3SStefano Zampini PetscFunctionReturn(0); 35d62866d3SStefano Zampini } 36d62866d3SStefano Zampini 37d62866d3SStefano Zampini #undef __FUNCT__ 38e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsCorrectionSolve" 39e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolve(PC pc, Vec rhs, Vec sol) 40e28d306cSStefano Zampini { 41e28d306cSStefano Zampini PetscErrorCode ierr; 42e28d306cSStefano Zampini 43e28d306cSStefano Zampini PetscFunctionBegin; 44e28d306cSStefano Zampini ierr = PCBDDCMumpsCorrectionSolve_Private(pc,rhs,sol,PETSC_FALSE);CHKERRQ(ierr); 45e28d306cSStefano Zampini PetscFunctionReturn(0); 46e28d306cSStefano Zampini } 47e28d306cSStefano Zampini 48e28d306cSStefano Zampini #undef __FUNCT__ 49e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsCorrectionSolveTranspose" 50e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolveTranspose(PC pc, Vec rhs, Vec sol) 51e28d306cSStefano Zampini { 52e28d306cSStefano Zampini PetscErrorCode ierr; 53e28d306cSStefano Zampini 54e28d306cSStefano Zampini PetscFunctionBegin; 55e28d306cSStefano Zampini ierr = PCBDDCMumpsCorrectionSolve_Private(pc,rhs,sol,PETSC_TRUE);CHKERRQ(ierr); 56e28d306cSStefano Zampini PetscFunctionReturn(0); 57e28d306cSStefano Zampini } 58e28d306cSStefano Zampini 59e28d306cSStefano Zampini #undef __FUNCT__ 60d62866d3SStefano Zampini #define __FUNCT__ "PCBDDCReuseMumpsReset" 61d62866d3SStefano Zampini static PetscErrorCode PCBDDCReuseMumpsReset(PCBDDCReuseMumps reuse) 62d62866d3SStefano Zampini { 63d62866d3SStefano Zampini PetscErrorCode ierr; 64d62866d3SStefano Zampini 65d62866d3SStefano Zampini PetscFunctionBegin; 66d62866d3SStefano Zampini ierr = MatDestroy(&reuse->F);CHKERRQ(ierr); 67d62866d3SStefano Zampini ierr = VecDestroy(&reuse->sol);CHKERRQ(ierr); 68d62866d3SStefano Zampini ierr = VecDestroy(&reuse->rhs);CHKERRQ(ierr); 69d62866d3SStefano Zampini ierr = PCDestroy(&reuse->interior_solver);CHKERRQ(ierr); 70d62866d3SStefano Zampini ierr = PCDestroy(&reuse->correction_solver);CHKERRQ(ierr); 7153892102SStefano Zampini ierr = ISDestroy(&reuse->is_R);CHKERRQ(ierr); 7253892102SStefano Zampini ierr = ISDestroy(&reuse->is_B);CHKERRQ(ierr); 73d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse->correction_scatter_B);CHKERRQ(ierr); 7453892102SStefano Zampini ierr = VecDestroy(&reuse->sol_B);CHKERRQ(ierr); 7553892102SStefano Zampini ierr = VecDestroy(&reuse->rhs_B);CHKERRQ(ierr); 76d62866d3SStefano Zampini PetscFunctionReturn(0); 77d62866d3SStefano Zampini } 78d5574798SStefano Zampini 79d5574798SStefano Zampini #undef __FUNCT__ 80e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolve_Private" 81e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose) 82d5574798SStefano Zampini { 83d62866d3SStefano Zampini PCBDDCReuseMumps ctx; 84d5574798SStefano Zampini PetscScalar *array,*array_mumps; 85ffbec71fSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 86d5574798SStefano Zampini PetscInt ival; 87ffbec71fSStefano Zampini #endif 88d5574798SStefano Zampini PetscErrorCode ierr; 89d5574798SStefano Zampini 90d5574798SStefano Zampini PetscFunctionBegin; 91d5574798SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 92d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 93d5574798SStefano Zampini ierr = MatMumpsGetIcntl(ctx->F,26,&ival);CHKERRQ(ierr); 94d5574798SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr); 95d62866d3SStefano Zampini #endif 96d5574798SStefano Zampini /* copy rhs into factored matrix workspace (can it be avoided?, MatSolve_MUMPS has another copy b->x internally) */ 97d5574798SStefano Zampini ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 98d5574798SStefano Zampini ierr = VecGetArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 99d62866d3SStefano Zampini ierr = PetscMemcpy(array_mumps,array,ctx->n_I*sizeof(PetscScalar));CHKERRQ(ierr); 100d5574798SStefano Zampini ierr = VecRestoreArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 101d5574798SStefano Zampini ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 102d5574798SStefano Zampini 103e28d306cSStefano Zampini if (transpose) { 104e28d306cSStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 105e28d306cSStefano Zampini } else { 106d5574798SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 107e28d306cSStefano Zampini } 108d5574798SStefano Zampini 109d5574798SStefano Zampini /* get back data to caller worskpace */ 110d5574798SStefano Zampini ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 111d5574798SStefano Zampini ierr = VecGetArray(sol,&array);CHKERRQ(ierr); 112d62866d3SStefano Zampini ierr = PetscMemcpy(array,array_mumps,ctx->n_I*sizeof(PetscScalar));CHKERRQ(ierr); 113d5574798SStefano Zampini ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr); 114d5574798SStefano Zampini ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 115d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 116d5574798SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,ival);CHKERRQ(ierr); 117d62866d3SStefano Zampini #endif 118d5574798SStefano Zampini PetscFunctionReturn(0); 119d5574798SStefano Zampini } 1203202ece2SStefano Zampini 1213202ece2SStefano Zampini #undef __FUNCT__ 122e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolve" 123e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC pc, Vec rhs, Vec sol) 124e28d306cSStefano Zampini { 125e28d306cSStefano Zampini PetscErrorCode ierr; 126e28d306cSStefano Zampini 127e28d306cSStefano Zampini PetscFunctionBegin; 128e28d306cSStefano Zampini ierr = PCBDDCMumpsInteriorSolve_Private(pc,rhs,sol,PETSC_FALSE);CHKERRQ(ierr); 129e28d306cSStefano Zampini PetscFunctionReturn(0); 130e28d306cSStefano Zampini } 131e28d306cSStefano Zampini 132e28d306cSStefano Zampini #undef __FUNCT__ 133e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolveTranspose" 134e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolveTranspose(PC pc, Vec rhs, Vec sol) 135e28d306cSStefano Zampini { 136e28d306cSStefano Zampini PetscErrorCode ierr; 137e28d306cSStefano Zampini 138e28d306cSStefano Zampini PetscFunctionBegin; 139e28d306cSStefano Zampini ierr = PCBDDCMumpsInteriorSolve_Private(pc,rhs,sol,PETSC_TRUE);CHKERRQ(ierr); 140e28d306cSStefano Zampini PetscFunctionReturn(0); 141e28d306cSStefano Zampini } 142e28d306cSStefano Zampini 143e28d306cSStefano Zampini #undef __FUNCT__ 1443202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur" 1455ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S) 1463202ece2SStefano Zampini { 1473202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 1483202ece2SStefano Zampini KSP ksp; 1493202ece2SStefano Zampini PC pc; 1503202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 1513202ece2SStefano Zampini PetscReal fill = 2.0; 152f11841e3SStefano Zampini PetscInt n_I; 1533202ece2SStefano Zampini PetscMPIInt size; 1543202ece2SStefano Zampini PetscErrorCode ierr; 1553202ece2SStefano Zampini 1563202ece2SStefano Zampini PetscFunctionBegin; 1573202ece2SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr); 1583202ece2SStefano Zampini if (size != 1) { 1593202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 1603202ece2SStefano Zampini } 161f11841e3SStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 162f11841e3SStefano Zampini PetscBool Sdense; 163f11841e3SStefano Zampini 164f11841e3SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr); 165f11841e3SStefano Zampini if (!Sdense) { 166f11841e3SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should dense"); 167f11841e3SStefano Zampini } 168f11841e3SStefano Zampini } 1693202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 1703202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 1713202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 1723202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 1733202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 1743202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 1753202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 1763202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 177f11841e3SStefano Zampini ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr); 178f11841e3SStefano Zampini if (n_I) { 1793202ece2SStefano Zampini if (!Bdense) { 1803202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 1813202ece2SStefano Zampini } else { 1823202ece2SStefano Zampini Bd = B; 1833202ece2SStefano Zampini } 1843202ece2SStefano Zampini 1853202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 1863202ece2SStefano Zampini Mat fact; 1873202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 1883202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 1893202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 1903202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 1913202ece2SStefano Zampini } else { 19207b1e237SStefano Zampini PetscBool ex = PETSC_TRUE; 19307b1e237SStefano Zampini 19407b1e237SStefano Zampini if (ex) { 1953202ece2SStefano Zampini Mat Ainvd; 1963202ece2SStefano Zampini 1973202ece2SStefano Zampini ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr); 1983202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 1993202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 20007b1e237SStefano Zampini } else { 20107b1e237SStefano Zampini Vec sol,rhs; 20207b1e237SStefano Zampini PetscScalar *arrayrhs,*arraysol; 20307b1e237SStefano Zampini PetscInt i,nrhs,n; 20407b1e237SStefano Zampini 20507b1e237SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 20607b1e237SStefano Zampini ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr); 20707b1e237SStefano Zampini ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr); 20807b1e237SStefano Zampini ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr); 20907b1e237SStefano Zampini ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr); 21007b1e237SStefano Zampini ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr); 21107b1e237SStefano Zampini for (i=0;i<nrhs;i++) { 21207b1e237SStefano Zampini ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr); 21307b1e237SStefano Zampini ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr); 21407b1e237SStefano Zampini ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr); 21507b1e237SStefano Zampini ierr = VecResetArray(rhs);CHKERRQ(ierr); 21607b1e237SStefano Zampini ierr = VecResetArray(sol);CHKERRQ(ierr); 21707b1e237SStefano Zampini } 21807b1e237SStefano Zampini ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr); 21907b1e237SStefano Zampini ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr); 22007b1e237SStefano Zampini } 2213202ece2SStefano Zampini } 2225ec10c6aSStefano Zampini if (!Bdense & !issym) { 2233202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 2243202ece2SStefano Zampini } 2255ec10c6aSStefano Zampini 2265ec10c6aSStefano Zampini if (!issym) { 2273202ece2SStefano Zampini if (!Cdense) { 2283202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 2293202ece2SStefano Zampini } else { 2303202ece2SStefano Zampini Cd = C; 2313202ece2SStefano Zampini } 2325ec10c6aSStefano Zampini ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 2333202ece2SStefano Zampini if (!Cdense) { 2343202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 2353202ece2SStefano Zampini } 2365ec10c6aSStefano Zampini } else { 2375ec10c6aSStefano Zampini ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 2385ec10c6aSStefano Zampini if (!Bdense) { 2395ec10c6aSStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 2405ec10c6aSStefano Zampini } 2415ec10c6aSStefano Zampini } 2425ec10c6aSStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 243f11841e3SStefano Zampini } 2443202ece2SStefano Zampini 2453202ece2SStefano Zampini if (D) { 2463202ece2SStefano Zampini Mat Dd; 2473202ece2SStefano Zampini PetscBool Ddense; 2483202ece2SStefano Zampini 2493202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 2503202ece2SStefano Zampini if (!Ddense) { 2513202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 2523202ece2SStefano Zampini } else { 2533202ece2SStefano Zampini Dd = D; 2543202ece2SStefano Zampini } 255f11841e3SStefano Zampini if (n_I) { 2563202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 257f11841e3SStefano Zampini } else { 258f11841e3SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 259f11841e3SStefano Zampini ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr); 260f11841e3SStefano Zampini } else { 261f11841e3SStefano Zampini ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 262f11841e3SStefano Zampini } 263f11841e3SStefano Zampini } 2643202ece2SStefano Zampini if (!Ddense) { 2653202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 2663202ece2SStefano Zampini } 2673202ece2SStefano Zampini } else { 2683202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 2693202ece2SStefano Zampini } 2703202ece2SStefano Zampini PetscFunctionReturn(0); 2713202ece2SStefano Zampini } 27234a97f8cSStefano Zampini 27334a97f8cSStefano Zampini #undef __FUNCT__ 2741580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp" 2756816873aSStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool faster_deluxe, PetscBool compute_Stilda, PetscBool reuse_solvers,PetscBool use_edges, PetscBool use_faces) 276b1b3d7a2SStefano Zampini { 277be83ff47SStefano Zampini Mat F,A_II,A_IB,A_BI,A_BB,AE_II; 278be83ff47SStefano Zampini Mat S_all; 279d2627357SStefano Zampini Mat global_schur_subsets,work_mat; 28008122e43SStefano Zampini Mat S_Ej_tilda_all,S_Ej_inv_all; 2815db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 282a9b99552SStefano Zampini IS is_I,is_I_layer,temp_is; 283d648f858SStefano Zampini PetscInt *nnz,*all_local_idx_G,*all_local_idx_N; 284eb595f79SStefano Zampini PetscInt *auxnum1,*auxnum2,*all_local_idx_G_rep; 2855a95e1ceSStefano Zampini PetscInt i,subset_size,max_subset_size; 286883469d8SStefano Zampini PetscInt extra,local_size,global_size; 28708122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 28806a4b1faSStefano Zampini PetscScalar *Bwork; 2895a95e1ceSStefano Zampini PetscSubcomm subcomm; 2905a95e1ceSStefano Zampini PetscMPIInt color,rank; 2915a95e1ceSStefano Zampini MPI_Comm comm_n; 292b1b3d7a2SStefano Zampini PetscErrorCode ierr; 293b1b3d7a2SStefano Zampini 294b1b3d7a2SStefano Zampini PetscFunctionBegin; 295a64f4aa4SStefano Zampini /* update info in sub_schurs */ 296a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 297a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 298a64f4aa4SStefano Zampini if (Ain) { 299a64f4aa4SStefano Zampini PetscBool isseqaij; 300a64f4aa4SStefano Zampini 301a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 302a64f4aa4SStefano Zampini if (isseqaij) { 303a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 304a64f4aa4SStefano Zampini sub_schurs->A = Ain; 305a64f4aa4SStefano Zampini } else { /* SeqBAIJ matrices does not support symmetry checking, SeqSBAIJ does not support MatPermute */ 306a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 307a64f4aa4SStefano Zampini } 308a64f4aa4SStefano Zampini } 309a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 310a64f4aa4SStefano Zampini sub_schurs->S = Sin; 311a64f4aa4SStefano Zampini if (sub_schurs->use_mumps) { 312a64f4aa4SStefano Zampini sub_schurs->use_mumps = (PetscBool)(!!sub_schurs->A); 313a64f4aa4SStefano Zampini } 314a64f4aa4SStefano Zampini 3155a95e1ceSStefano Zampini /* preliminary checks */ 3165a95e1ceSStefano Zampini if (!sub_schurs->use_mumps && compute_Stilda) { 3175a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3185a95e1ceSStefano Zampini } 3195a95e1ceSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 3205a95e1ceSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 3215a95e1ceSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 3225a95e1ceSStefano Zampini if (sub_schurs->A) { 3235a95e1ceSStefano Zampini PetscInt lsize; 3245a95e1ceSStefano Zampini 3255a95e1ceSStefano Zampini ierr = MatGetSize(sub_schurs->A,&lsize,NULL);CHKERRQ(ierr); 3265a95e1ceSStefano Zampini if (lsize) { 3275a95e1ceSStefano Zampini ierr = MatIsHermitian(sub_schurs->A,0.0,&sub_schurs->is_hermitian);CHKERRQ(ierr); 3285a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 3295a95e1ceSStefano Zampini PetscScalar val; 3305a95e1ceSStefano Zampini Vec vec1,vec2; 3315a95e1ceSStefano Zampini 3325a95e1ceSStefano Zampini ierr = MatCreateVecs(sub_schurs->A,&vec1,&vec2);CHKERRQ(ierr); 3335a95e1ceSStefano Zampini ierr = VecSetRandom(vec1,NULL); 3345a95e1ceSStefano Zampini ierr = VecCopy(vec1,vec2);CHKERRQ(ierr); 3355a95e1ceSStefano Zampini ierr = MatMult(sub_schurs->A,vec2,vec1);CHKERRQ(ierr); 3365a95e1ceSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 337c62d2063SStefano Zampini if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE; 3385a95e1ceSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 3395a95e1ceSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 3405a95e1ceSStefano Zampini } 3415a95e1ceSStefano Zampini } else { 3425a95e1ceSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 3435a95e1ceSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 3445a95e1ceSStefano Zampini } 3455a95e1ceSStefano Zampini if (compute_Stilda && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 346c62d2063SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"General matrix pencils are not currently supported (%d,%d)",sub_schurs->is_hermitian,sub_schurs->is_posdef); 3475a95e1ceSStefano Zampini } 3485a95e1ceSStefano Zampini } 3495a95e1ceSStefano Zampini /* restrict work on active processes */ 3505a95e1ceSStefano Zampini color = 0; 3515a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 3525a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 3535a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 3545a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 3555a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3565a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 3575a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3585a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 3595a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 3605a95e1ceSStefano Zampini PetscFunctionReturn(0); 3615a95e1ceSStefano Zampini } 3625a95e1ceSStefano Zampini 363b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 364883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 365a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 366f11841e3SStefano Zampini PetscBool isseqaij; 367a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 368a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 369f11841e3SStefano Zampini if (!isseqaij) { 370a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 371a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 372a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 373a64f4aa4SStefano Zampini } else { 374a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 375a64f4aa4SStefano Zampini A_BB = tA_BB; 376a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 377a64f4aa4SStefano Zampini A_IB = tA_IB; 378a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 379a64f4aa4SStefano Zampini A_BI = tA_BI; 380f11841e3SStefano Zampini } 381a58a30b4SStefano Zampini } else { 3825a95e1ceSStefano Zampini A_II = NULL; 3835a95e1ceSStefano Zampini A_IB = NULL; 3845a95e1ceSStefano Zampini A_BI = NULL; 3855a95e1ceSStefano Zampini A_BB = NULL; 386b1b3d7a2SStefano Zampini } 3875a95e1ceSStefano Zampini S_all = NULL; 3885a95e1ceSStefano Zampini S_Ej_tilda_all = NULL; 3895a95e1ceSStefano Zampini S_Ej_inv_all = NULL; 390b1b3d7a2SStefano Zampini 391b1b3d7a2SStefano Zampini /* determine interior problems */ 3923dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 3933dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 394b1b3d7a2SStefano Zampini PetscBT touched; 395b1b3d7a2SStefano Zampini const PetscInt* idx_B; 396b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 397b1b3d7a2SStefano Zampini 3983dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 3993dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 4003dc780c3SStefano Zampini } 401b1b3d7a2SStefano Zampini /* get sizes */ 402b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 403b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 404b1b3d7a2SStefano Zampini 405b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 406b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 407b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 408b1b3d7a2SStefano Zampini 409b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 410b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 411b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 412b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 413b1b3d7a2SStefano Zampini } 414b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 415b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 416b1b3d7a2SStefano Zampini 417b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 418b1b3d7a2SStefano Zampini n_local_dofs = n_B; 419b1b3d7a2SStefano Zampini n_prev_added = n_B; 420b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 421b1b3d7a2SStefano Zampini PetscInt n_added; 422b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 423b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 424b1b3d7a2SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+n_B); 425b1b3d7a2SStefano Zampini } 426b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 427b1b3d7a2SStefano Zampini n_prev_added = n_added; 428b1b3d7a2SStefano Zampini n_local_dofs += n_added; 429b1b3d7a2SStefano Zampini if (!n_added) break; 430b1b3d7a2SStefano Zampini } 431b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 432b1b3d7a2SStefano Zampini 433883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 434a9b99552SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&is_I_layer);CHKERRQ(ierr); 435b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 436a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 437883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 438883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 439b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 440b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 441a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 442b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 443b1b3d7a2SStefano Zampini 444b1b3d7a2SStefano Zampini /* II block */ 445b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 446b1b3d7a2SStefano Zampini } 447b1b3d7a2SStefano Zampini } else { 448b1b3d7a2SStefano Zampini PetscInt n_I; 449b1b3d7a2SStefano Zampini 450b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 451b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 452a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 453b1b3d7a2SStefano Zampini 454b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 455883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 456b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 457b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 458b1b3d7a2SStefano Zampini 459b1b3d7a2SStefano Zampini /* II block is the same */ 460b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 461b1b3d7a2SStefano Zampini AE_II = A_II; 462b1b3d7a2SStefano Zampini } 463b1b3d7a2SStefano Zampini } 4645a95e1ceSStefano Zampini 465883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 4665a95e1ceSStefano Zampini max_subset_size = 0; 467883469d8SStefano Zampini local_size = 0; 4685a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4695a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4705a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 471883469d8SStefano Zampini local_size += subset_size; 472883469d8SStefano Zampini } 473883469d8SStefano Zampini 474883469d8SStefano Zampini /* Work arrays for local indices */ 475883469d8SStefano Zampini extra = 0; 476883469d8SStefano Zampini if (sub_schurs->use_mumps) { 477a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 478883469d8SStefano Zampini } 479883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 480883469d8SStefano Zampini if (extra) { 481883469d8SStefano Zampini const PetscInt *idxs; 482a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 483883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 484a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 485883469d8SStefano Zampini } 486883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 487eb595f79SStefano Zampini ierr = PetscMalloc2(sub_schurs->n_subs,&auxnum1,sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 488883469d8SStefano Zampini 489883469d8SStefano Zampini /* Get local indices in local numbering */ 490883469d8SStefano Zampini local_size = 0; 4915a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 492883469d8SStefano Zampini PetscInt j; 493883469d8SStefano Zampini const PetscInt *idxs; 494883469d8SStefano Zampini 4955a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4965a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 497eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 498eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 499eb595f79SStefano Zampini auxnum2[i] = subset_size; 500883469d8SStefano Zampini /* subset indices in local numbering */ 501883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 5025a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 503883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 504883469d8SStefano Zampini local_size += subset_size; 505883469d8SStefano Zampini } 506883469d8SStefano Zampini 5075a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 508d2627357SStefano Zampini Bwork = NULL; 509d2627357SStefano Zampini pivots = NULL; 5105a95e1ceSStefano Zampini if (sub_schurs->n_subs && !sub_schurs->is_hermitian) { 511d2627357SStefano Zampini PetscScalar lwork; 512d2627357SStefano Zampini 513d2627357SStefano Zampini B_lwork = -1; 514d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 515d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 516d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 517d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 518d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 519d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 520d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 521d2627357SStefano Zampini } 522d2627357SStefano Zampini 523d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 524eb595f79SStefano Zampini ierr = PCBDDCSubsetNumbering(comm_n,sub_schurs->l2gmap,sub_schurs->n_subs,auxnum1,auxnum2,&global_size,&all_local_idx_G_rep);CHKERRQ(ierr); 525eb595f79SStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_G);CHKERRQ(ierr); 526eb595f79SStefano Zampini local_size = 0; 527eb595f79SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 528eb595f79SStefano Zampini PetscInt j; 529eb595f79SStefano Zampini for (j=0;j<auxnum2[i];j++) all_local_idx_G[local_size++] = all_local_idx_G_rep[i] + j; 530eb595f79SStefano Zampini } 531eb595f79SStefano Zampini ierr = PetscFree(all_local_idx_G_rep);CHKERRQ(ierr); 532eb595f79SStefano Zampini ierr = PetscFree2(auxnum1,auxnum2);CHKERRQ(ierr); 5335a95e1ceSStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,local_size,all_local_idx_G,PETSC_COPY_VALUES,&l2gmap_subsets);CHKERRQ(ierr); 5345a95e1ceSStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 535d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 536d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 537d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 538d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 5392972d61bSStefano Zampini 5405a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 5415a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 5425a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 5435a95e1ceSStefano Zampini 5445a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 5455a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 5465a95e1ceSStefano Zampini if (subset_size != local_size) { 5475a95e1ceSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 5485a95e1ceSStefano Zampini } 5495a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 550b1b3d7a2SStefano Zampini } 551b1b3d7a2SStefano Zampini 5525a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 5535a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 5545a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 5555a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 5565a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 5575a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 558aa83b6aeSStefano Zampini } else { 5595a95e1ceSStefano Zampini ierr = MatZeroEntries(sub_schurs->S_Ej_all);CHKERRQ(ierr); 560aa83b6aeSStefano Zampini } 561b1b3d7a2SStefano Zampini 5625a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 563be83ff47SStefano Zampini F = NULL; 5645a95e1ceSStefano Zampini ierr = PetscBTMemzero(sub_schurs->n_subs,sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 5655a95e1ceSStefano Zampini if (!sub_schurs->use_mumps) { 5665a95e1ceSStefano Zampini Mat S_Ej_expl; 5675a95e1ceSStefano Zampini PetscScalar *work; 5685a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 5695a95e1ceSStefano Zampini PetscBool Sdense; 5705a95e1ceSStefano Zampini 5715a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 5725a95e1ceSStefano Zampini local_size = 0; 573b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5745a95e1ceSStefano Zampini IS is_subset_B; 5755a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 5765a95e1ceSStefano Zampini 5775a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 5785a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 5795a95e1ceSStefano Zampini /* EE block */ 5805a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 5815a95e1ceSStefano Zampini /* IE block */ 5825a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 5835a95e1ceSStefano Zampini /* EI block */ 5845a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 5855a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 5865a95e1ceSStefano Zampini } else { 5875a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 5885a95e1ceSStefano Zampini } 589a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 5905a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 5915a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 5925a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 5935a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 594b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 595b1b3d7a2SStefano Zampini KSP ksp; 596b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 5975a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 598b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 599b1b3d7a2SStefano Zampini KSP origksp,schurksp; 600b1b3d7a2SStefano Zampini PC origpc,schurpc; 601b1b3d7a2SStefano Zampini KSPType ksp_type; 602b1b3d7a2SStefano Zampini PetscInt n_internal; 6035a95e1ceSStefano Zampini PetscBool ispcnone; 604b1b3d7a2SStefano Zampini 605b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 6065a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 607b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 608b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 609b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 610b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 6115a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 6125a95e1ceSStefano Zampini if (!ispcnone) { 6135a95e1ceSStefano Zampini PCType pc_type; 614b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 615b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 6165a95e1ceSStefano Zampini } else { 6175a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 6185a95e1ceSStefano Zampini } 619b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 620b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 621b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 622b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 623b1b3d7a2SStefano Zampini if (solver) { 624b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 625b1b3d7a2SStefano Zampini } 626b1b3d7a2SStefano Zampini } 627b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 628b1b3d7a2SStefano Zampini } 6295a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6305a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 6315a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 6325a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 6335a95e1ceSStefano Zampini if (Sdense) { 6345a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 6355a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 636b1b3d7a2SStefano Zampini } 6375a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 6385a95e1ceSStefano Zampini } else { 6395a95e1ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 6405a95e1ceSStefano Zampini } 6415a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 642a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 6435a95e1ceSStefano Zampini local_size += subset_size; 6445a95e1ceSStefano Zampini } 6455a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 646b1b3d7a2SStefano Zampini /* free */ 647b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 648b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 6495a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 650883469d8SStefano Zampini } else { 651be83ff47SStefano Zampini Mat A; 652883469d8SStefano Zampini IS is_A_all; 653be83ff47SStefano Zampini PetscScalar *work,*S_data; 654be83ff47SStefano Zampini PetscInt *idxs_schur,n_I,n_I_all,*dummy_idx,size_schur,size_active_schur,cum,cum2; 655*4a6c6b0dSStefano Zampini PetscBool mumps_S; 656883469d8SStefano Zampini 657883469d8SStefano Zampini /* get working mat */ 658a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 659d62866d3SStefano Zampini if (!sub_schurs->is_dir) { 660883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 661d62866d3SStefano Zampini size_schur = local_size; 662d62866d3SStefano Zampini } else { 663d62866d3SStefano Zampini IS list[2]; 664d62866d3SStefano Zampini 665d62866d3SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&list[0]);CHKERRQ(ierr); 666d62866d3SStefano Zampini list[1] = sub_schurs->is_dir; 667d62866d3SStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&is_A_all);CHKERRQ(ierr); 668d62866d3SStefano Zampini ierr = ISDestroy(&list[0]);CHKERRQ(ierr); 669d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&size_schur);CHKERRQ(ierr); 670d62866d3SStefano Zampini size_schur += local_size; 671d62866d3SStefano Zampini } 67253892102SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 67353892102SStefano Zampini size_active_schur = local_size; /* size active schurs does not count any dirichlet dof on the interface */ 6746816873aSStefano Zampini ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 675a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 676883469d8SStefano Zampini 67708122e43SStefano Zampini if (n_I) { 6789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 679d62866d3SStefano Zampini ierr = MatSetOption(A,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 680883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 681883469d8SStefano Zampini } else { 682d62866d3SStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr); 683883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 684883469d8SStefano Zampini } 685883469d8SStefano Zampini 686883469d8SStefano Zampini /* subsets ordered last */ 687d62866d3SStefano Zampini ierr = PetscMalloc1(size_schur,&idxs_schur);CHKERRQ(ierr); 688d62866d3SStefano Zampini for (i=0;i<size_schur;i++) { 689883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 690883469d8SStefano Zampini } 6915a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 692d62866d3SStefano Zampini ierr = MatMumpsSetSchurIndices(F,size_schur,idxs_schur);CHKERRQ(ierr); 6935a95e1ceSStefano Zampini #endif 694883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 695883469d8SStefano Zampini 696883469d8SStefano Zampini /* factorization step */ 6979ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 698883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 699be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 700be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 701be83ff47SStefano Zampini #endif 702883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 703883469d8SStefano Zampini } else { 704883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 705be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 706be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 707be83ff47SStefano Zampini #endif 708883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 709883469d8SStefano Zampini } 710883469d8SStefano Zampini 711883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 7125a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 713883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 7145a95e1ceSStefano Zampini #endif 715d5574798SStefano Zampini 716d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 717d5574798SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I_all);CHKERRQ(ierr); 718be83ff47SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && (n_I == n_I_all)); 719*4a6c6b0dSStefano Zampini mumps_S = PETSC_TRUE; 720be83ff47SStefano Zampini } else { /* we can't use MUMPS when size_schur == size_of_the_problem */ 721be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 722be83ff47SStefano Zampini reuse_solvers = PETSC_FALSE; 723*4a6c6b0dSStefano Zampini mumps_S = PETSC_FALSE; 724be83ff47SStefano Zampini } 725be83ff47SStefano Zampini 726be83ff47SStefano Zampini if (reuse_solvers) { 727d62866d3SStefano Zampini Mat A_II; 72853892102SStefano Zampini Vec vec1_B; 729d62866d3SStefano Zampini PCBDDCReuseMumps msolv_ctx; 730d5574798SStefano Zampini 731d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 7326816873aSStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 733e28d306cSStefano Zampini } else { 734e28d306cSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_mumps);CHKERRQ(ierr); 735d62866d3SStefano Zampini } 736e28d306cSStefano Zampini msolv_ctx = sub_schurs->reuse_mumps; 737be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 738be83ff47SStefano Zampini ierr = MatGetSize(A_II,&msolv_ctx->n_I,NULL);CHKERRQ(ierr); 739d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 740d5574798SStefano Zampini msolv_ctx->F = F; 741d5574798SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,&msolv_ctx->rhs);CHKERRQ(ierr); 742d62866d3SStefano Zampini 743d62866d3SStefano Zampini /* interior solver */ 744d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->interior_solver);CHKERRQ(ierr); 745d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 746d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 747d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 748d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolve);CHKERRQ(ierr); 749e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolveTranspose);CHKERRQ(ierr); 750d62866d3SStefano Zampini 751d62866d3SStefano Zampini /* correction solver */ 752d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->correction_solver);CHKERRQ(ierr); 753d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,A,A);CHKERRQ(ierr); 754d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 755d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 756d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolve);CHKERRQ(ierr); 757e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolveTranspose);CHKERRQ(ierr); 75853892102SStefano Zampini 75953892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 76053892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 76153892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 76253892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 76353892102SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 76453892102SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 76553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 76653892102SStefano Zampini msolv_ctx->is_R = is_A_all; 767d5574798SStefano Zampini } 76808122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 76953892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 7705db18549SStefano Zampini 771be83ff47SStefano Zampini /* Work arrays */ 772be83ff47SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 77312d906b1SStefano Zampini if (compute_Stilda) { 774a1337663SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_tilda_all);CHKERRQ(ierr); 775d62866d3SStefano Zampini ierr = MatSetSizes(S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 776a1337663SStefano Zampini ierr = MatSetType(S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 777a1337663SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 77808122e43SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&S_Ej_inv_all);CHKERRQ(ierr); 779d62866d3SStefano Zampini ierr = MatSetSizes(S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 78008122e43SStefano Zampini ierr = MatSetType(S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 78108122e43SStefano Zampini ierr = MatSeqAIJSetPreallocation(S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 78212d906b1SStefano Zampini } 783d2627357SStefano Zampini 784be83ff47SStefano Zampini /* S_Ej_all */ 785be83ff47SStefano Zampini cum = cum2 = 0; 786be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 78765d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 78865d8bf0aSStefano Zampini PetscInt j; 78965d8bf0aSStefano Zampini 7905a95e1ceSStefano Zampini /* get S_E */ 791b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 792*4a6c6b0dSStefano Zampini if (mumps_S && sub_schurs->is_hermitian) { /* When using MUMPS data I need to expand to upper triangular (column oriented) */ 793be83ff47SStefano Zampini PetscInt k; 794be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 795be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 796be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 797be83ff47SStefano Zampini work[j*subset_size+k] = S_data[cum2+k*size_schur+j]; 798be83ff47SStefano Zampini } 799be83ff47SStefano Zampini } 800be83ff47SStefano Zampini } else { /* copy to workspace */ 801be83ff47SStefano Zampini PetscInt k; 802be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 803be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 804be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 805be83ff47SStefano Zampini } 806be83ff47SStefano Zampini } 8079087bf02SStefano Zampini } 8085a95e1ceSStefano Zampini /* insert S_E values */ 809be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 8105a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 811a1337663SStefano Zampini 812be83ff47SStefano Zampini /* if adaptivity is requested, invert S_E block */ 813d2627357SStefano Zampini if (compute_Stilda && ((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || (!PetscBTLookup(sub_schurs->is_edge,i) && use_faces))) { 81408122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 81508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 816be83ff47SStefano Zampini if (!sub_schurs->is_hermitian) { /* TODO add sytrf/i for symmetric non hermitian */ 8175a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 81808122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 8195a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 82008122e43SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 8212972d61bSStefano Zampini } else { 8225a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 8232972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 8245a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 8252972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 8262972d61bSStefano Zampini } 82708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 8285a95e1ceSStefano Zampini ierr = MatSetValues(S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8299087bf02SStefano Zampini } 830be83ff47SStefano Zampini cum += subset_size; 831be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 832be83ff47SStefano Zampini } 833be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 834*4a6c6b0dSStefano Zampini 835be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 836*4a6c6b0dSStefano Zampini if (mumps_S) { 837be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 838*4a6c6b0dSStefano Zampini } 839*4a6c6b0dSStefano Zampini #endif 840*4a6c6b0dSStefano Zampini 841*4a6c6b0dSStefano Zampini if (compute_Stilda) { 842*4a6c6b0dSStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur) { /* we already computed the inverse */ 843*4a6c6b0dSStefano Zampini PetscInt j; 844*4a6c6b0dSStefano Zampini for (j=0;j<size_schur;j++) dummy_idx[j] = j; 845*4a6c6b0dSStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 846*4a6c6b0dSStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,0);CHKERRQ(ierr); 847*4a6c6b0dSStefano Zampini } else { 848*4a6c6b0dSStefano Zampini if (mumps_S) { /* use MatMumps calls to invert S */ 849*4a6c6b0dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 850be83ff47SStefano Zampini ierr = MatMumpsInvertSchurComplement(F);CHKERRQ(ierr); 851be83ff47SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 852*4a6c6b0dSStefano Zampini #endif 853*4a6c6b0dSStefano Zampini } else { /* we need to invert explicitly since we are not using MUMPS for S */ 854be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 855be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 856be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 857be83ff47SStefano Zampini if (!sub_schurs->is_hermitian) { /* TODO add sytrf/i for symmetric non hermitian */ 858be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 859be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 860be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 861be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 862be83ff47SStefano Zampini } else { 863be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 864be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 865be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 866be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 867be83ff47SStefano Zampini } 868be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 869be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 870be83ff47SStefano Zampini } 871be83ff47SStefano Zampini /* S_Ej_tilda_all */ 872be83ff47SStefano Zampini cum = cum2 = 0; 873be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 874be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 875be83ff47SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 876be83ff47SStefano Zampini /* if adaptivity is requested, invert S_E and insert St_E^-1 blocks */ 877*4a6c6b0dSStefano Zampini if (((PetscBTLookup(sub_schurs->is_edge,i) && use_edges) || (!PetscBTLookup(sub_schurs->is_edge,i) && use_faces))) { 878be83ff47SStefano Zampini PetscInt j; 879be83ff47SStefano Zampini /* get (St^-1)_E */ 880be83ff47SStefano Zampini if (sub_schurs->is_hermitian) { /* I need to expand to upper triangular (column oriented) */ 881be83ff47SStefano Zampini PetscInt k; 882be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 883be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 884be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 885be83ff47SStefano Zampini work[j*subset_size+k] = S_data[cum2+k*size_schur+j]; 886be83ff47SStefano Zampini } 887be83ff47SStefano Zampini } 888be83ff47SStefano Zampini } else { /* copy to workspace */ 889be83ff47SStefano Zampini PetscInt k; 890be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 891be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 892be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 893be83ff47SStefano Zampini } 894be83ff47SStefano Zampini } 895be83ff47SStefano Zampini } 896be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 8975a95e1ceSStefano Zampini ierr = MatSetValues(S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8989087bf02SStefano Zampini ierr = PetscBTSet(sub_schurs->computed_Stilda_subs,i);CHKERRQ(ierr); 89908122e43SStefano Zampini } 900be83ff47SStefano Zampini cum += subset_size; 901be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 902883469d8SStefano Zampini } 903be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 904be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 905*4a6c6b0dSStefano Zampini if (mumps_S) { 906be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 9075db18549SStefano Zampini } 908be83ff47SStefano Zampini #endif 9099087bf02SStefano Zampini } 910*4a6c6b0dSStefano Zampini } 911*4a6c6b0dSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 912*4a6c6b0dSStefano Zampini } 913be83ff47SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 914a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 915a1337663SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 916a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 917a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 918a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 919a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 9205db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9215db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9225a95e1ceSStefano Zampini if (compute_Stilda) { 923a1337663SStefano Zampini ierr = MatAssemblyBegin(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 924a1337663SStefano Zampini ierr = MatAssemblyEnd(S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 92508122e43SStefano Zampini ierr = MatAssemblyBegin(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 92608122e43SStefano Zampini ierr = MatAssemblyEnd(S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 92708122e43SStefano Zampini } 928a1337663SStefano Zampini 9295db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 9305db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 9313927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 9323927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 9335a95e1ceSStefano Zampini 9345db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 9355a95e1ceSStefano Zampini ierr = ISCreateGeneral(comm_n,local_size,all_local_idx_G,PETSC_OWN_POINTER,&temp_is);CHKERRQ(ierr); 936d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 937d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 93808122e43SStefano Zampini 939ac632422SStefano Zampini /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */ 940ac632422SStefano Zampini if (faster_deluxe) { 9415a95e1ceSStefano Zampini Mat tmpmat; 9425a95e1ceSStefano Zampini PetscScalar *array; 9435a95e1ceSStefano Zampini PetscInt cum; 9445a95e1ceSStefano Zampini 9455a95e1ceSStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 9465a95e1ceSStefano Zampini cum = 0; 9475a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 9485a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 9495a95e1ceSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 9505a95e1ceSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9515a95e1ceSStefano Zampini if (!sub_schurs->is_hermitian) { 9525a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 9535a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 9545a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 9555a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 9565a95e1ceSStefano Zampini } else { 9575a95e1ceSStefano Zampini PetscInt j,k; 9585a95e1ceSStefano Zampini 9595a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 9605a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 9615a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 9625a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 9635a95e1ceSStefano Zampini for (j=0;j<B_N;j++) { 9645a95e1ceSStefano Zampini for (k=j+1;k<B_N;k++) { 9655a95e1ceSStefano Zampini array[k*B_N+j+cum] = array[j*B_N+k+cum]; 9665a95e1ceSStefano Zampini } 9675a95e1ceSStefano Zampini } 9685a95e1ceSStefano Zampini } 9695a95e1ceSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 9705a95e1ceSStefano Zampini cum += subset_size*subset_size; 9715a95e1ceSStefano Zampini } 9725a95e1ceSStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 9735a95e1ceSStefano Zampini ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr); 9745a95e1ceSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 975ac632422SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 9765a95e1ceSStefano Zampini sub_schurs->S_Ej_all = tmpmat; 9775a95e1ceSStefano Zampini } 9785a95e1ceSStefano Zampini 979f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 9805a95e1ceSStefano Zampini if (compute_Stilda) { 981a1337663SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_tilda_all);CHKERRQ(ierr); 982a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 983d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 984d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 98508122e43SStefano Zampini ierr = MatISSetLocalMat(work_mat,S_Ej_inv_all);CHKERRQ(ierr); 98608122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 987d648f858SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 988d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(global_schur_subsets,temp_is,temp_is,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 98908122e43SStefano Zampini } 9903202ece2SStefano Zampini 9915a95e1ceSStefano Zampini /* free workspace */ 99206a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 993a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 994a1337663SStefano Zampini ierr = MatDestroy(&S_Ej_tilda_all);CHKERRQ(ierr); 99508122e43SStefano Zampini ierr = MatDestroy(&S_Ej_inv_all);CHKERRQ(ierr); 9963202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 9975db18549SStefano Zampini ierr = ISDestroy(&temp_is);CHKERRQ(ierr); 9985a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 999b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1000b1b3d7a2SStefano Zampini } 1001b1b3d7a2SStefano Zampini 1002b1b3d7a2SStefano Zampini #undef __FUNCT__ 1003b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 1004a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 1005b1b3d7a2SStefano Zampini { 10069bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 10075a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1008b1b3d7a2SStefano Zampini PetscBool is_sorted; 1009b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1010b1b3d7a2SStefano Zampini 1011b1b3d7a2SStefano Zampini PetscFunctionBegin; 1012b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 1013b1b3d7a2SStefano Zampini if (!is_sorted) { 1014b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1015b1b3d7a2SStefano Zampini } 1016b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 1017b1b3d7a2SStefano Zampini if (!is_sorted) { 1018b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1019b1b3d7a2SStefano Zampini } 1020b1b3d7a2SStefano Zampini 1021b1b3d7a2SStefano Zampini /* reset any previous data */ 1022b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1023b1b3d7a2SStefano Zampini 10245a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 10259bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1026b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 102708122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 102808122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 1029b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1030b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 1031b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1032b1b3d7a2SStefano Zampini } 1033b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 1034b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 103508122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1036b1b3d7a2SStefano Zampini } 1037b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 1038b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 1039d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1040d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1041b1b3d7a2SStefano Zampini 10426816873aSStefano Zampini /* Determine if MUMPS can be used */ 1043883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 1044883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1045a64f4aa4SStefano Zampini sub_schurs->use_mumps = PETSC_TRUE; 1046883469d8SStefano Zampini #endif 1047b1b3d7a2SStefano Zampini 1048b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1049b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1050b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1051b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 10525db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 10535db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 10545db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 10555db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 10565a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1057b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1058a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { /* sort by local ordering mumps is not present */ 1059b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1060b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 1061b96c3477SStefano Zampini } 10629bb4a8caSStefano Zampini } 1063d62866d3SStefano Zampini sub_schurs->is_vertices = vertices; 1064b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1065b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 106608122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1067b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1068b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1069b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1070b1b3d7a2SStefano Zampini } 1071b1b3d7a2SStefano Zampini 1072b1b3d7a2SStefano Zampini #undef __FUNCT__ 107334a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 107434a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 107534a97f8cSStefano Zampini { 107634a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 107734a97f8cSStefano Zampini PetscErrorCode ierr; 107834a97f8cSStefano Zampini 107934a97f8cSStefano Zampini PetscFunctionBegin; 108034a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 10815ff63025SStefano Zampini schurs_ctx->n_subs = 0; 108234a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 108334a97f8cSStefano Zampini PetscFunctionReturn(0); 108434a97f8cSStefano Zampini } 108534a97f8cSStefano Zampini 108634a97f8cSStefano Zampini #undef __FUNCT__ 108734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 108834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 108934a97f8cSStefano Zampini { 109034a97f8cSStefano Zampini PetscErrorCode ierr; 109134a97f8cSStefano Zampini 109234a97f8cSStefano Zampini PetscFunctionBegin; 109334a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 109434a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 109534a97f8cSStefano Zampini PetscFunctionReturn(0); 109634a97f8cSStefano Zampini } 109734a97f8cSStefano Zampini 109834a97f8cSStefano Zampini #undef __FUNCT__ 109934a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 110034a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 110134a97f8cSStefano Zampini { 110234a97f8cSStefano Zampini PetscInt i; 110334a97f8cSStefano Zampini PetscErrorCode ierr; 110434a97f8cSStefano Zampini 110534a97f8cSStefano Zampini PetscFunctionBegin; 11061e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1107b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1108b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1109b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 11105db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 11115db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 111241c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 111341c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 111408122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1115a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 11165db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 1117d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 1118d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 111908122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 112008122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->computed_Stilda_subs);CHKERRQ(ierr); 112134a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1122b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 112334a97f8cSStefano Zampini } 11245ff63025SStefano Zampini if (sub_schurs->n_subs) { 1125b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 11263dc780c3SStefano Zampini } 1127d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1128d62866d3SStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 1129d62866d3SStefano Zampini } 1130d62866d3SStefano Zampini ierr = PetscFree(sub_schurs->reuse_mumps);CHKERRQ(ierr); 113134a97f8cSStefano Zampini sub_schurs->n_subs = 0; 113234a97f8cSStefano Zampini PetscFunctionReturn(0); 113334a97f8cSStefano Zampini } 113434a97f8cSStefano Zampini 113534a97f8cSStefano Zampini #undef __FUNCT__ 113634a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 11372a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 113834a97f8cSStefano Zampini { 113934a97f8cSStefano Zampini PetscInt i,j,n; 114034a97f8cSStefano Zampini PetscErrorCode ierr; 114134a97f8cSStefano Zampini 114234a97f8cSStefano Zampini PetscFunctionBegin; 114334a97f8cSStefano Zampini n = 0; 114434a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 114534a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 114634a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 114734a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 114834a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 114934a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 115034a97f8cSStefano Zampini queue_tip[n] = dof; 115134a97f8cSStefano Zampini n++; 115234a97f8cSStefano Zampini } 115334a97f8cSStefano Zampini } 115434a97f8cSStefano Zampini } 115534a97f8cSStefano Zampini *n_added = n; 115634a97f8cSStefano Zampini PetscFunctionReturn(0); 115734a97f8cSStefano Zampini } 1158