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" 275862806e4SStefano 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) 276b1b3d7a2SStefano Zampini { 277be83ff47SStefano Zampini Mat F,A_II,A_IB,A_BI,A_BB,AE_II; 278be83ff47SStefano Zampini Mat S_all; 2795a94c880SStefano Zampini Mat global_schur_subsets,work_mat,*submats; 2805db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 281*dc456d91SStefano Zampini IS is_I,is_I_layer; 282*dc456d91SStefano Zampini IS all_subsets,all_subsets_mult,all_subsets_n; 283*dc456d91SStefano Zampini PetscInt *nnz,*all_local_idx_N; 284*dc456d91SStefano Zampini PetscInt *auxnum1,*auxnum2; 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); 2983301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 2993301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 300a64f4aa4SStefano Zampini if (Ain) { 301a64f4aa4SStefano Zampini PetscBool isseqaij; 3023301b35fSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 3033301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3043301b35fSStefano Zampini if (Ain->symmetric_set) { 3053301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->symmetric; 3063301b35fSStefano Zampini } 3073301b35fSStefano Zampini #else 3083301b35fSStefano Zampini if (Ain->hermitian_set) { 3093301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->hermitian; 3103301b35fSStefano Zampini } 3113301b35fSStefano Zampini #endif 3123301b35fSStefano Zampini if (Ain->spd_set) { 3133301b35fSStefano Zampini sub_schurs->is_posdef = Ain->spd; 3143301b35fSStefano Zampini } 315a64f4aa4SStefano Zampini 3163301b35fSStefano Zampini /* check */ 317a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3183301b35fSStefano Zampini if (compute_Stilda && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 3193301b35fSStefano Zampini PetscInt lsize; 3203301b35fSStefano Zampini 3213301b35fSStefano Zampini ierr = MatGetSize(Ain,&lsize,NULL);CHKERRQ(ierr); 3223301b35fSStefano Zampini if (lsize) { 3233301b35fSStefano Zampini PetscScalar val; 3243301b35fSStefano Zampini PetscReal norm; 3253301b35fSStefano Zampini Vec vec1,vec2,vec3; 3263301b35fSStefano Zampini 3273301b35fSStefano Zampini ierr = MatCreateVecs(Ain,&vec1,&vec2);CHKERRQ(ierr); 3283301b35fSStefano Zampini ierr = VecDuplicate(vec1,&vec3);CHKERRQ(ierr); 3293301b35fSStefano Zampini ierr = VecSetRandom(vec1,NULL);CHKERRQ(ierr); 3303301b35fSStefano Zampini ierr = MatMult(Ain,vec1,vec2);CHKERRQ(ierr); 3313301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3323301b35fSStefano Zampini ierr = MatMultTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 3333301b35fSStefano Zampini #else 3343301b35fSStefano Zampini ierr = MatMultHermitianTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 3353301b35fSStefano Zampini #endif 3363301b35fSStefano Zampini ierr = VecAXPY(vec3,-1.0,vec2);CHKERRQ(ierr); 3373301b35fSStefano Zampini ierr = VecNorm(vec3,NORM_INFINITY,&norm);CHKERRQ(ierr); 338862806e4SStefano Zampini if (norm > PetscSqrtReal(PETSC_SMALL)) { 3393301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 3403301b35fSStefano Zampini } else { 3413301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 3423301b35fSStefano Zampini } 3433301b35fSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 3443301b35fSStefano Zampini if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE; 3453301b35fSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 3463301b35fSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 3473301b35fSStefano Zampini ierr = VecDestroy(&vec3);CHKERRQ(ierr); 3483301b35fSStefano Zampini } else { 3493301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 3503301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 3513301b35fSStefano Zampini } 3523301b35fSStefano Zampini } 353a64f4aa4SStefano Zampini if (isseqaij) { 354a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 355a64f4aa4SStefano Zampini sub_schurs->A = Ain; 3563301b35fSStefano Zampini } else { 357a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 358a64f4aa4SStefano Zampini } 359a64f4aa4SStefano Zampini } 3603301b35fSStefano Zampini if (compute_Stilda && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 3613301b35fSStefano 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); 3623301b35fSStefano Zampini } 3633301b35fSStefano Zampini 364a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 365a64f4aa4SStefano Zampini sub_schurs->S = Sin; 366a64f4aa4SStefano Zampini if (sub_schurs->use_mumps) { 367a64f4aa4SStefano Zampini sub_schurs->use_mumps = (PetscBool)(!!sub_schurs->A); 368a64f4aa4SStefano Zampini } 369a64f4aa4SStefano Zampini 3705a95e1ceSStefano Zampini /* preliminary checks */ 3715a95e1ceSStefano Zampini if (!sub_schurs->use_mumps && compute_Stilda) { 3725a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3735a95e1ceSStefano Zampini } 3745a95e1ceSStefano Zampini 3755a95e1ceSStefano Zampini /* restrict work on active processes */ 3765a95e1ceSStefano Zampini color = 0; 3775a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 3785a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 3795a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 3805a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 3815a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3825a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 3835a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3845a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 3855a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 3865a95e1ceSStefano Zampini PetscFunctionReturn(0); 3875a95e1ceSStefano Zampini } 3885a95e1ceSStefano Zampini 389b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 390883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 391a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 3923301b35fSStefano Zampini PetscBool isseqsbaij; 393a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 3943301b35fSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr); 3953301b35fSStefano Zampini if (isseqsbaij) { 396a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 397a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 398a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 399a64f4aa4SStefano Zampini } else { 400a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 401a64f4aa4SStefano Zampini A_BB = tA_BB; 402a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 403a64f4aa4SStefano Zampini A_IB = tA_IB; 404a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 405a64f4aa4SStefano Zampini A_BI = tA_BI; 406f11841e3SStefano Zampini } 407a58a30b4SStefano Zampini } else { 4085a95e1ceSStefano Zampini A_II = NULL; 4095a95e1ceSStefano Zampini A_IB = NULL; 4105a95e1ceSStefano Zampini A_BI = NULL; 4115a95e1ceSStefano Zampini A_BB = NULL; 412b1b3d7a2SStefano Zampini } 4135a95e1ceSStefano Zampini S_all = NULL; 414b1b3d7a2SStefano Zampini 415b1b3d7a2SStefano Zampini /* determine interior problems */ 4163dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 4173dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 418b1b3d7a2SStefano Zampini PetscBT touched; 419b1b3d7a2SStefano Zampini const PetscInt* idx_B; 420b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 421b1b3d7a2SStefano Zampini 4223dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 4233dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 4243dc780c3SStefano Zampini } 425b1b3d7a2SStefano Zampini /* get sizes */ 426b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 427b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 428b1b3d7a2SStefano Zampini 429b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 430b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 431b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 432b1b3d7a2SStefano Zampini 433b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 434b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 435b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 436b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 437b1b3d7a2SStefano Zampini } 438b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 439b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 440b1b3d7a2SStefano Zampini 441b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 442b1b3d7a2SStefano Zampini n_local_dofs = n_B; 443b1b3d7a2SStefano Zampini n_prev_added = n_B; 444b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 445b1b3d7a2SStefano Zampini PetscInt n_added; 446b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 447b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 448b1b3d7a2SStefano 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); 449b1b3d7a2SStefano Zampini } 450b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 451b1b3d7a2SStefano Zampini n_prev_added = n_added; 452b1b3d7a2SStefano Zampini n_local_dofs += n_added; 453b1b3d7a2SStefano Zampini if (!n_added) break; 454b1b3d7a2SStefano Zampini } 455b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 456b1b3d7a2SStefano Zampini 457883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 458a9b99552SStefano 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); 459b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 460a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 461883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 462883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 463b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 464b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 465a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 466b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 467b1b3d7a2SStefano Zampini 468b1b3d7a2SStefano Zampini /* II block */ 469b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 470b1b3d7a2SStefano Zampini } 471b1b3d7a2SStefano Zampini } else { 472b1b3d7a2SStefano Zampini PetscInt n_I; 473b1b3d7a2SStefano Zampini 474b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 475b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 476a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 477b1b3d7a2SStefano Zampini 478b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 479883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 480b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 481b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 482b1b3d7a2SStefano Zampini 483b1b3d7a2SStefano Zampini /* II block is the same */ 484b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 485b1b3d7a2SStefano Zampini AE_II = A_II; 486b1b3d7a2SStefano Zampini } 487b1b3d7a2SStefano Zampini } 4885a95e1ceSStefano Zampini 489883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 4905a95e1ceSStefano Zampini max_subset_size = 0; 491883469d8SStefano Zampini local_size = 0; 4925a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4935a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4945a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 495883469d8SStefano Zampini local_size += subset_size; 496883469d8SStefano Zampini } 497883469d8SStefano Zampini 498883469d8SStefano Zampini /* Work arrays for local indices */ 499883469d8SStefano Zampini extra = 0; 500883469d8SStefano Zampini if (sub_schurs->use_mumps) { 501a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 502883469d8SStefano Zampini } 503883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 504883469d8SStefano Zampini if (extra) { 505883469d8SStefano Zampini const PetscInt *idxs; 506a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 507883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 508a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 509883469d8SStefano Zampini } 510883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 511*dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr); 512*dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 513883469d8SStefano Zampini 514883469d8SStefano Zampini /* Get local indices in local numbering */ 515883469d8SStefano Zampini local_size = 0; 5165a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 517883469d8SStefano Zampini PetscInt j; 518883469d8SStefano Zampini const PetscInt *idxs; 519883469d8SStefano Zampini 5205a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5215a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 522eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 523eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 524eb595f79SStefano Zampini auxnum2[i] = subset_size; 525883469d8SStefano Zampini /* subset indices in local numbering */ 526883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 5275a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 528883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 529883469d8SStefano Zampini local_size += subset_size; 530883469d8SStefano Zampini } 531883469d8SStefano Zampini 5325a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 533d2627357SStefano Zampini Bwork = NULL; 534d2627357SStefano Zampini pivots = NULL; 535d6462365SStefano Zampini if (!sub_schurs->is_hermitian || !sub_schurs->is_posdef) { 536d2627357SStefano Zampini PetscScalar lwork; 537d2627357SStefano Zampini 538d2627357SStefano Zampini B_lwork = -1; 539d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 540d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 541d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 542d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 543d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 544d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 545d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 546d2627357SStefano Zampini } 547d2627357SStefano Zampini 548d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 549*dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr); 550*dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr); 551*dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 552*dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr); 553*dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr); 554*dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets);CHKERRQ(ierr); 555*dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr); 556*dc456d91SStefano Zampini ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr); 557*dc456d91SStefano Zampini if (i != local_size) { 558*dc456d91SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %d != %d",i,local_size); 559eb595f79SStefano Zampini } 560*dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(all_subsets_n,&l2gmap_subsets);CHKERRQ(ierr); 5615a95e1ceSStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,&work_mat);CHKERRQ(ierr); 562d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 563d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 564d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 565d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 5662972d61bSStefano Zampini 5675a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 5685a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 5695a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 5705a95e1ceSStefano Zampini 5715a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 5725a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 5735a95e1ceSStefano Zampini if (subset_size != local_size) { 5745a95e1ceSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 5755a95e1ceSStefano Zampini } 5765a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 577b1b3d7a2SStefano Zampini } 578b1b3d7a2SStefano Zampini 5795a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 5805a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 5815a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 5825a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 5835a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 5845a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 585aa83b6aeSStefano Zampini } 586b1b3d7a2SStefano Zampini 5875a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 588be83ff47SStefano Zampini F = NULL; 5895a95e1ceSStefano Zampini if (!sub_schurs->use_mumps) { 5905a95e1ceSStefano Zampini Mat S_Ej_expl; 5915a95e1ceSStefano Zampini PetscScalar *work; 5925a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 5935a95e1ceSStefano Zampini PetscBool Sdense; 5945a95e1ceSStefano Zampini 5955a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 5965a95e1ceSStefano Zampini local_size = 0; 597b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5985a95e1ceSStefano Zampini IS is_subset_B; 5995a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 6005a95e1ceSStefano Zampini 6015a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 6025a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 6035a95e1ceSStefano Zampini /* EE block */ 6045a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 6055a95e1ceSStefano Zampini /* IE block */ 6065a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 6075a95e1ceSStefano Zampini /* EI block */ 6085a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 6095a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 6105a95e1ceSStefano Zampini } else { 6115a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 6125a95e1ceSStefano Zampini } 613a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 6145a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 6155a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 6165a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 6175a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 618b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 619b1b3d7a2SStefano Zampini KSP ksp; 620b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 6215a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 622b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 623b1b3d7a2SStefano Zampini KSP origksp,schurksp; 624b1b3d7a2SStefano Zampini PC origpc,schurpc; 625b1b3d7a2SStefano Zampini KSPType ksp_type; 626b1b3d7a2SStefano Zampini PetscInt n_internal; 6275a95e1ceSStefano Zampini PetscBool ispcnone; 628b1b3d7a2SStefano Zampini 629b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 6305a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 631b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 632b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 633b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 634b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 6355a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 6365a95e1ceSStefano Zampini if (!ispcnone) { 6375a95e1ceSStefano Zampini PCType pc_type; 638b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 639b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 6405a95e1ceSStefano Zampini } else { 6415a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 6425a95e1ceSStefano Zampini } 643b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 644b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 645b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 646b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 647b1b3d7a2SStefano Zampini if (solver) { 648b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 649b1b3d7a2SStefano Zampini } 650b1b3d7a2SStefano Zampini } 651b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 652b1b3d7a2SStefano Zampini } 6535a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6545a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 6555a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 6565a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 6575a95e1ceSStefano Zampini if (Sdense) { 6585a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 6595a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 660b1b3d7a2SStefano Zampini } 6615a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 6625a95e1ceSStefano Zampini } else { 6635a95e1ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 6645a95e1ceSStefano Zampini } 6655a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 666a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 6675a95e1ceSStefano Zampini local_size += subset_size; 6685a95e1ceSStefano Zampini } 6695a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 670b1b3d7a2SStefano Zampini /* free */ 671b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 672b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 6735a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 674883469d8SStefano Zampini } else { 675be83ff47SStefano Zampini Mat A; 676883469d8SStefano Zampini IS is_A_all; 677be83ff47SStefano Zampini PetscScalar *work,*S_data; 678be83ff47SStefano Zampini PetscInt *idxs_schur,n_I,n_I_all,*dummy_idx,size_schur,size_active_schur,cum,cum2; 6794a6c6b0dSStefano Zampini PetscBool mumps_S; 680883469d8SStefano Zampini 681883469d8SStefano Zampini /* get working mat */ 682a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 683d62866d3SStefano Zampini if (!sub_schurs->is_dir) { 684883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 685d62866d3SStefano Zampini size_schur = local_size; 686d62866d3SStefano Zampini } else { 687d62866d3SStefano Zampini IS list[2]; 688d62866d3SStefano Zampini 689d62866d3SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&list[0]);CHKERRQ(ierr); 690d62866d3SStefano Zampini list[1] = sub_schurs->is_dir; 691d62866d3SStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&is_A_all);CHKERRQ(ierr); 692d62866d3SStefano Zampini ierr = ISDestroy(&list[0]);CHKERRQ(ierr); 693d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&size_schur);CHKERRQ(ierr); 694d62866d3SStefano Zampini size_schur += local_size; 695d62866d3SStefano Zampini } 69653892102SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 69753892102SStefano Zampini size_active_schur = local_size; /* size active schurs does not count any dirichlet dof on the interface */ 6986816873aSStefano Zampini ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 699a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 7003301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr); 7013301b35fSStefano Zampini ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 7023301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 703883469d8SStefano Zampini 70408122e43SStefano Zampini if (n_I) { 7059ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 706883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 707883469d8SStefano Zampini } else { 708883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 709883469d8SStefano Zampini } 710883469d8SStefano Zampini /* subsets ordered last */ 711d62866d3SStefano Zampini ierr = PetscMalloc1(size_schur,&idxs_schur);CHKERRQ(ierr); 712d62866d3SStefano Zampini for (i=0;i<size_schur;i++) { 713883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 714883469d8SStefano Zampini } 7155a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 716d62866d3SStefano Zampini ierr = MatMumpsSetSchurIndices(F,size_schur,idxs_schur);CHKERRQ(ierr); 7175a95e1ceSStefano Zampini #endif 718883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 719883469d8SStefano Zampini 720883469d8SStefano Zampini /* factorization step */ 7219ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 722883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 723be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 724be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 725be83ff47SStefano Zampini #endif 726883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 727883469d8SStefano Zampini } else { 728883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 729be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 730be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 731be83ff47SStefano Zampini #endif 732883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 733883469d8SStefano Zampini } 734883469d8SStefano Zampini 735883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 7365a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 737883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 7385a95e1ceSStefano Zampini #endif 739d5574798SStefano Zampini 740d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 741d5574798SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I_all);CHKERRQ(ierr); 742be83ff47SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && (n_I == n_I_all)); 7434a6c6b0dSStefano Zampini mumps_S = PETSC_TRUE; 744be83ff47SStefano Zampini } else { /* we can't use MUMPS when size_schur == size_of_the_problem */ 745be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 746be83ff47SStefano Zampini reuse_solvers = PETSC_FALSE; 7474a6c6b0dSStefano Zampini mumps_S = PETSC_FALSE; 748be83ff47SStefano Zampini } 749be83ff47SStefano Zampini 750be83ff47SStefano Zampini if (reuse_solvers) { 751d62866d3SStefano Zampini Mat A_II; 75253892102SStefano Zampini Vec vec1_B; 753d62866d3SStefano Zampini PCBDDCReuseMumps msolv_ctx; 754d5574798SStefano Zampini 755d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 7566816873aSStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 757e28d306cSStefano Zampini } else { 758e28d306cSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_mumps);CHKERRQ(ierr); 759d62866d3SStefano Zampini } 760e28d306cSStefano Zampini msolv_ctx = sub_schurs->reuse_mumps; 761be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 762be83ff47SStefano Zampini ierr = MatGetSize(A_II,&msolv_ctx->n_I,NULL);CHKERRQ(ierr); 763d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 764d5574798SStefano Zampini msolv_ctx->F = F; 765d5574798SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,&msolv_ctx->rhs);CHKERRQ(ierr); 766d62866d3SStefano Zampini 767d62866d3SStefano Zampini /* interior solver */ 768d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->interior_solver);CHKERRQ(ierr); 769d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 770d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 771d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 772d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolve);CHKERRQ(ierr); 773e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolveTranspose);CHKERRQ(ierr); 774d62866d3SStefano Zampini 775d62866d3SStefano Zampini /* correction solver */ 776d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->correction_solver);CHKERRQ(ierr); 777d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,A,A);CHKERRQ(ierr); 778d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 779d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 780d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolve);CHKERRQ(ierr); 781e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolveTranspose);CHKERRQ(ierr); 78253892102SStefano Zampini 78353892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 78453892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 78553892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 78653892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 78753892102SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 78853892102SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 78953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 79053892102SStefano Zampini msolv_ctx->is_R = is_A_all; 791d5574798SStefano Zampini } 79208122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 79353892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 7945db18549SStefano Zampini 795be83ff47SStefano Zampini /* Work arrays */ 796be83ff47SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 7975a94c880SStefano Zampini 7985a94c880SStefano Zampini /* matrices for adaptive selection */ 79912d906b1SStefano Zampini if (compute_Stilda) { 8005a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_tilda_all) { 8015a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 8025a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 8035a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 8045a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 8055a94c880SStefano Zampini } 8065a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_inv_all) { 8075a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 8085a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 8095a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 8105a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 8115a94c880SStefano Zampini } 81212d906b1SStefano Zampini } 813d2627357SStefano Zampini 814be83ff47SStefano Zampini /* S_Ej_all */ 815be83ff47SStefano Zampini cum = cum2 = 0; 816be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 81765d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 81865d8bf0aSStefano Zampini PetscInt j; 81965d8bf0aSStefano Zampini 8205a95e1ceSStefano Zampini /* get S_E */ 821b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 8224a6c6b0dSStefano Zampini if (mumps_S && sub_schurs->is_hermitian) { /* When using MUMPS data I need to expand to upper triangular (column oriented) */ 823be83ff47SStefano Zampini PetscInt k; 824be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 825be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 826be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 827be83ff47SStefano Zampini work[j*subset_size+k] = S_data[cum2+k*size_schur+j]; 828be83ff47SStefano Zampini } 829be83ff47SStefano Zampini } 830be83ff47SStefano Zampini } else { /* copy to workspace */ 831be83ff47SStefano Zampini PetscInt k; 832be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 833be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 834be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 835be83ff47SStefano Zampini } 836be83ff47SStefano Zampini } 8379087bf02SStefano Zampini } 8385a95e1ceSStefano Zampini /* insert S_E values */ 839be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 8405a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 841a1337663SStefano Zampini 842be83ff47SStefano Zampini /* if adaptivity is requested, invert S_E block */ 843862806e4SStefano Zampini if (compute_Stilda) { 84408122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 84508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 846d6462365SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { /* TODO add sytrf/i for symmetric non hermitian */ 8475a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 8482972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 8495a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 8502972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 851d6462365SStefano Zampini } else { 852d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 853d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 854d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 855d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 8562972d61bSStefano Zampini } 85708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 8585a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8599087bf02SStefano Zampini } 860be83ff47SStefano Zampini cum += subset_size; 861be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 862be83ff47SStefano Zampini } 863be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 8644a6c6b0dSStefano Zampini 865be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 8664a6c6b0dSStefano Zampini if (mumps_S) { 867be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 8684a6c6b0dSStefano Zampini } 8694a6c6b0dSStefano Zampini #endif 8704a6c6b0dSStefano Zampini 8714a6c6b0dSStefano Zampini if (compute_Stilda) { 8724a6c6b0dSStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur) { /* we already computed the inverse */ 8734a6c6b0dSStefano Zampini PetscInt j; 8744a6c6b0dSStefano Zampini for (j=0;j<size_schur;j++) dummy_idx[j] = j; 8755a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8764a6c6b0dSStefano Zampini } else { 8774a6c6b0dSStefano Zampini if (mumps_S) { /* use MatMumps calls to invert S */ 8784a6c6b0dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 879be83ff47SStefano Zampini ierr = MatMumpsInvertSchurComplement(F);CHKERRQ(ierr); 880be83ff47SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 8814a6c6b0dSStefano Zampini #endif 8824a6c6b0dSStefano Zampini } else { /* we need to invert explicitly since we are not using MUMPS for S */ 883be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 884be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 885be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 886d6462365SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { /* TODO add sytrf/i for symmetric non hermitian */ 887be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 888be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 889be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 890be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 891d6462365SStefano Zampini } else { 892d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 893d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 894d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 895d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 896be83ff47SStefano Zampini } 897be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 898be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 899be83ff47SStefano Zampini } 900be83ff47SStefano Zampini /* S_Ej_tilda_all */ 901be83ff47SStefano Zampini cum = cum2 = 0; 902be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 903be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 904be83ff47SStefano Zampini PetscInt j; 905862806e4SStefano Zampini 906862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 907be83ff47SStefano Zampini /* get (St^-1)_E */ 908aff50787SStefano Zampini if (sub_schurs->is_hermitian) { /* Here I don't need to expand to upper triangular (column oriented) */ 909be83ff47SStefano Zampini PetscInt k; 910be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 911be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 912be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 913be83ff47SStefano Zampini } 914be83ff47SStefano Zampini } 915be83ff47SStefano Zampini } else { /* copy to workspace */ 916be83ff47SStefano Zampini PetscInt k; 917be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 918be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 919be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 920be83ff47SStefano Zampini } 921be83ff47SStefano Zampini } 922be83ff47SStefano Zampini } 923be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 9245a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 925be83ff47SStefano Zampini cum += subset_size; 926be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 927883469d8SStefano Zampini } 928be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 929be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 9304a6c6b0dSStefano Zampini if (mumps_S) { 931be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 9325db18549SStefano Zampini } 933be83ff47SStefano Zampini #endif 9349087bf02SStefano Zampini } 9354a6c6b0dSStefano Zampini } 9364a6c6b0dSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 9374a6c6b0dSStefano Zampini } 9385a94c880SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 939be83ff47SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 940a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 941a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 942a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 943a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 944a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 9455db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9465db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9475a95e1ceSStefano Zampini if (compute_Stilda) { 9485a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9495a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9505a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9515a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 95208122e43SStefano Zampini } 953a1337663SStefano Zampini 9545db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 9555db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 9563927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 9573927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 9585a95e1ceSStefano Zampini 9595db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 9605a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_all) { 961*dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_INITIAL_MATRIX,&submats);CHKERRQ(ierr); 9625a94c880SStefano Zampini sub_schurs->sum_S_Ej_all = submats[0]; 9635a94c880SStefano Zampini } else { 9645a94c880SStefano Zampini ierr = PetscMalloc1(1,&submats);CHKERRQ(ierr); 9655a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_all; 966*dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 9675a94c880SStefano Zampini } 96808122e43SStefano Zampini 969ac632422SStefano Zampini /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */ 970ac632422SStefano Zampini if (faster_deluxe) { 9715a95e1ceSStefano Zampini Mat tmpmat; 9725a95e1ceSStefano Zampini PetscScalar *array; 9735a95e1ceSStefano Zampini PetscInt cum; 9745a95e1ceSStefano Zampini 9755a95e1ceSStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 9765a95e1ceSStefano Zampini cum = 0; 9775a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 9785a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 9795a95e1ceSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 9805a95e1ceSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 981d6462365SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 9825a95e1ceSStefano Zampini PetscInt j,k; 9835a95e1ceSStefano Zampini 9845a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 9855a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 9865a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 9875a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 9885a95e1ceSStefano Zampini for (j=0;j<B_N;j++) { 9895a95e1ceSStefano Zampini for (k=j+1;k<B_N;k++) { 9905a95e1ceSStefano Zampini array[k*B_N+j+cum] = array[j*B_N+k+cum]; 9915a95e1ceSStefano Zampini } 9925a95e1ceSStefano Zampini } 993d6462365SStefano Zampini } else { 994d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 995d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 996d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 997d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 9985a95e1ceSStefano Zampini } 9995a95e1ceSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10005a95e1ceSStefano Zampini cum += subset_size*subset_size; 10015a95e1ceSStefano Zampini } 10025a95e1ceSStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 10035a95e1ceSStefano Zampini ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr); 10045a95e1ceSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 1005ac632422SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 10065a95e1ceSStefano Zampini sub_schurs->S_Ej_all = tmpmat; 10075a95e1ceSStefano Zampini } 10085a95e1ceSStefano Zampini 1009f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 10105a95e1ceSStefano Zampini if (compute_Stilda) { 10115a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 1012a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 10135a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_tilda_all; 1014*dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 10155a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 101608122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 10175a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_inv_all; 1018*dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 101908122e43SStefano Zampini } 10203202ece2SStefano Zampini 10215a95e1ceSStefano Zampini /* free workspace */ 10225a94c880SStefano Zampini ierr = PetscFree(submats);CHKERRQ(ierr); 102306a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 1024a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 10253202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1026*dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 10275a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 1028b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1029b1b3d7a2SStefano Zampini } 1030b1b3d7a2SStefano Zampini 1031b1b3d7a2SStefano Zampini #undef __FUNCT__ 1032b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 1033a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 1034b1b3d7a2SStefano Zampini { 10359bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 10365a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1037b1b3d7a2SStefano Zampini PetscBool is_sorted; 1038b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1039b1b3d7a2SStefano Zampini 1040b1b3d7a2SStefano Zampini PetscFunctionBegin; 1041b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 1042b1b3d7a2SStefano Zampini if (!is_sorted) { 1043b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1044b1b3d7a2SStefano Zampini } 1045b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 1046b1b3d7a2SStefano Zampini if (!is_sorted) { 1047b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1048b1b3d7a2SStefano Zampini } 1049b1b3d7a2SStefano Zampini 1050b1b3d7a2SStefano Zampini /* reset any previous data */ 1051b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1052b1b3d7a2SStefano Zampini 10535a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 10549bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1055b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 105608122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 1057b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1058b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 1059b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1060b1b3d7a2SStefano Zampini } 1061b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 1062b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 106308122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1064b1b3d7a2SStefano Zampini } 1065b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 1066b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 1067d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1068d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1069b1b3d7a2SStefano Zampini 10706816873aSStefano Zampini /* Determine if MUMPS can be used */ 1071883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 1072883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1073a64f4aa4SStefano Zampini sub_schurs->use_mumps = PETSC_TRUE; 1074883469d8SStefano Zampini #endif 1075b1b3d7a2SStefano Zampini 1076b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1077b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1078b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1079b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 10805db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 10815db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 10825db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 10835db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 10845a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1085b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1086a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { /* sort by local ordering mumps is not present */ 1087b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1088b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 1089b96c3477SStefano Zampini } 10909bb4a8caSStefano Zampini } 1091d62866d3SStefano Zampini sub_schurs->is_vertices = vertices; 1092b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1093b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 109408122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1095b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1096b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1097b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1098b1b3d7a2SStefano Zampini } 1099b1b3d7a2SStefano Zampini 1100b1b3d7a2SStefano Zampini #undef __FUNCT__ 110134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 110234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 110334a97f8cSStefano Zampini { 110434a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 110534a97f8cSStefano Zampini PetscErrorCode ierr; 110634a97f8cSStefano Zampini 110734a97f8cSStefano Zampini PetscFunctionBegin; 110834a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 11095ff63025SStefano Zampini schurs_ctx->n_subs = 0; 111034a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 111134a97f8cSStefano Zampini PetscFunctionReturn(0); 111234a97f8cSStefano Zampini } 111334a97f8cSStefano Zampini 111434a97f8cSStefano Zampini #undef __FUNCT__ 111534a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 111634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 111734a97f8cSStefano Zampini { 111834a97f8cSStefano Zampini PetscErrorCode ierr; 111934a97f8cSStefano Zampini 112034a97f8cSStefano Zampini PetscFunctionBegin; 112134a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 112234a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 112334a97f8cSStefano Zampini PetscFunctionReturn(0); 112434a97f8cSStefano Zampini } 112534a97f8cSStefano Zampini 112634a97f8cSStefano Zampini #undef __FUNCT__ 112734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 112834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 112934a97f8cSStefano Zampini { 113034a97f8cSStefano Zampini PetscInt i; 113134a97f8cSStefano Zampini PetscErrorCode ierr; 113234a97f8cSStefano Zampini 113334a97f8cSStefano Zampini PetscFunctionBegin; 11341e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1135b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1136b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1137b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 11385db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 11395db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 114041c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 114141c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 114208122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1143a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 11445db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 1145d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 1146d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 114708122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 114834a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1149b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 115034a97f8cSStefano Zampini } 11515ff63025SStefano Zampini if (sub_schurs->n_subs) { 1152b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 11533dc780c3SStefano Zampini } 1154d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1155d62866d3SStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 1156d62866d3SStefano Zampini } 1157d62866d3SStefano Zampini ierr = PetscFree(sub_schurs->reuse_mumps);CHKERRQ(ierr); 115834a97f8cSStefano Zampini sub_schurs->n_subs = 0; 115934a97f8cSStefano Zampini PetscFunctionReturn(0); 116034a97f8cSStefano Zampini } 116134a97f8cSStefano Zampini 116234a97f8cSStefano Zampini #undef __FUNCT__ 116334a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 11642a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 116534a97f8cSStefano Zampini { 116634a97f8cSStefano Zampini PetscInt i,j,n; 116734a97f8cSStefano Zampini PetscErrorCode ierr; 116834a97f8cSStefano Zampini 116934a97f8cSStefano Zampini PetscFunctionBegin; 117034a97f8cSStefano Zampini n = 0; 117134a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 117234a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 117334a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 117434a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 117534a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 117634a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 117734a97f8cSStefano Zampini queue_tip[n] = dof; 117834a97f8cSStefano Zampini n++; 117934a97f8cSStefano Zampini } 118034a97f8cSStefano Zampini } 118134a97f8cSStefano Zampini } 118234a97f8cSStefano Zampini *n_added = n; 118334a97f8cSStefano Zampini PetscFunctionReturn(0); 118434a97f8cSStefano Zampini } 1185