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" 27506a4e24aSStefano 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 benign_trick) 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; 281dc456d91SStefano Zampini IS is_I,is_I_layer; 282dc456d91SStefano Zampini IS all_subsets,all_subsets_mult,all_subsets_n; 283dc456d91SStefano Zampini PetscInt *nnz,*all_local_idx_N; 284dc456d91SStefano 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; 29206a4e24aSStefano Zampini PetscBool use_getr = PETSC_FALSE; 293b1b3d7a2SStefano Zampini PetscErrorCode ierr; 294b1b3d7a2SStefano Zampini 295b1b3d7a2SStefano Zampini PetscFunctionBegin; 296a64f4aa4SStefano Zampini /* update info in sub_schurs */ 297a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 298a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 2993301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 3003301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 301a64f4aa4SStefano Zampini if (Ain) { 302a64f4aa4SStefano Zampini PetscBool isseqaij; 3033301b35fSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 3043301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3053301b35fSStefano Zampini if (Ain->symmetric_set) { 3063301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->symmetric; 3073301b35fSStefano Zampini } 3083301b35fSStefano Zampini #else 3093301b35fSStefano Zampini if (Ain->hermitian_set) { 3103301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->hermitian; 3113301b35fSStefano Zampini } 3123301b35fSStefano Zampini #endif 3133301b35fSStefano Zampini if (Ain->spd_set) { 3143301b35fSStefano Zampini sub_schurs->is_posdef = Ain->spd; 3153301b35fSStefano Zampini } 316a64f4aa4SStefano Zampini 3173301b35fSStefano Zampini /* check */ 318a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3191dac5d76SStefano Zampini if (compute_Stilda && (!Ain->hermitian_set || !Ain->spd_set)) { /* these are lazy checks, maybe I should throw an error if they are not set */ 3203301b35fSStefano Zampini PetscInt lsize; 3213301b35fSStefano Zampini 3223301b35fSStefano Zampini ierr = MatGetSize(Ain,&lsize,NULL);CHKERRQ(ierr); 3233301b35fSStefano Zampini if (lsize) { 3243301b35fSStefano Zampini PetscScalar val; 3253301b35fSStefano Zampini PetscReal norm; 3263301b35fSStefano Zampini Vec vec1,vec2,vec3; 3273301b35fSStefano Zampini 3283301b35fSStefano Zampini ierr = MatCreateVecs(Ain,&vec1,&vec2);CHKERRQ(ierr); 3293301b35fSStefano Zampini ierr = VecDuplicate(vec1,&vec3);CHKERRQ(ierr); 3303301b35fSStefano Zampini ierr = VecSetRandom(vec1,NULL);CHKERRQ(ierr); 3313301b35fSStefano Zampini ierr = MatMult(Ain,vec1,vec2);CHKERRQ(ierr); 3323301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3333301b35fSStefano Zampini ierr = MatMultTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 3343301b35fSStefano Zampini #else 3353301b35fSStefano Zampini ierr = MatMultHermitianTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 3363301b35fSStefano Zampini #endif 3373301b35fSStefano Zampini ierr = VecAXPY(vec3,-1.0,vec2);CHKERRQ(ierr); 3383301b35fSStefano Zampini ierr = VecNorm(vec3,NORM_INFINITY,&norm);CHKERRQ(ierr); 3391dac5d76SStefano Zampini if (!Ain->hermitian_set) { 340862806e4SStefano Zampini if (norm > PetscSqrtReal(PETSC_SMALL)) { 3413301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 3423301b35fSStefano Zampini } else { 3433301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 3443301b35fSStefano Zampini } 3451dac5d76SStefano Zampini } 3461dac5d76SStefano Zampini if (!Ain->spd_set && !benign_trick) { 3473301b35fSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 3483301b35fSStefano Zampini if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE; 34906a4e24aSStefano Zampini } 3503301b35fSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 3513301b35fSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 3523301b35fSStefano Zampini ierr = VecDestroy(&vec3);CHKERRQ(ierr); 3533301b35fSStefano Zampini } else { 3543301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 3553301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 3563301b35fSStefano Zampini } 3573301b35fSStefano Zampini } 358a64f4aa4SStefano Zampini if (isseqaij) { 359a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 360a64f4aa4SStefano Zampini sub_schurs->A = Ain; 3613301b35fSStefano Zampini } else { 362a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 363a64f4aa4SStefano Zampini } 364a64f4aa4SStefano Zampini } 3653301b35fSStefano Zampini 366a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 367a64f4aa4SStefano Zampini sub_schurs->S = Sin; 368a64f4aa4SStefano Zampini if (sub_schurs->use_mumps) { 369a64f4aa4SStefano Zampini sub_schurs->use_mumps = (PetscBool)(!!sub_schurs->A); 370a64f4aa4SStefano Zampini } 371a64f4aa4SStefano Zampini 3725a95e1ceSStefano Zampini /* preliminary checks */ 3735a95e1ceSStefano Zampini if (!sub_schurs->use_mumps && compute_Stilda) { 3745a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3755a95e1ceSStefano Zampini } 3765a95e1ceSStefano Zampini 3775a95e1ceSStefano Zampini /* restrict work on active processes */ 3785a95e1ceSStefano Zampini color = 0; 3795a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 3805a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 3815a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 3825a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 3835a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3845a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 3855a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3865a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 3875a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 3885a95e1ceSStefano Zampini PetscFunctionReturn(0); 3895a95e1ceSStefano Zampini } 39081ea8064SStefano Zampini /* ierr = PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL);CHKERRQ(ierr); */ 3915a95e1ceSStefano Zampini 392b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 393883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 394a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 3953301b35fSStefano Zampini PetscBool isseqsbaij; 396a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 3973301b35fSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr); 3983301b35fSStefano Zampini if (isseqsbaij) { 399a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 400a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 401a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 402a64f4aa4SStefano Zampini } else { 403a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 404a64f4aa4SStefano Zampini A_BB = tA_BB; 405a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 406a64f4aa4SStefano Zampini A_IB = tA_IB; 407a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 408a64f4aa4SStefano Zampini A_BI = tA_BI; 409f11841e3SStefano Zampini } 410a58a30b4SStefano Zampini } else { 4115a95e1ceSStefano Zampini A_II = NULL; 4125a95e1ceSStefano Zampini A_IB = NULL; 4135a95e1ceSStefano Zampini A_BI = NULL; 4145a95e1ceSStefano Zampini A_BB = NULL; 415b1b3d7a2SStefano Zampini } 4165a95e1ceSStefano Zampini S_all = NULL; 417b1b3d7a2SStefano Zampini 418b1b3d7a2SStefano Zampini /* determine interior problems */ 4193dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 4203dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 421b1b3d7a2SStefano Zampini PetscBT touched; 422b1b3d7a2SStefano Zampini const PetscInt* idx_B; 423b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 424b1b3d7a2SStefano Zampini 4253dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 4263dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 4273dc780c3SStefano Zampini } 428b1b3d7a2SStefano Zampini /* get sizes */ 429b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 430b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 431b1b3d7a2SStefano Zampini 432b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 433b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 434b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 435b1b3d7a2SStefano Zampini 436b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 437b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 438b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 439b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 440b1b3d7a2SStefano Zampini } 441b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 442b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 443b1b3d7a2SStefano Zampini 444b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 445b1b3d7a2SStefano Zampini n_local_dofs = n_B; 446b1b3d7a2SStefano Zampini n_prev_added = n_B; 447b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 448b1b3d7a2SStefano Zampini PetscInt n_added; 449b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 450b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 451b1b3d7a2SStefano 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); 452b1b3d7a2SStefano Zampini } 453b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 454b1b3d7a2SStefano Zampini n_prev_added = n_added; 455b1b3d7a2SStefano Zampini n_local_dofs += n_added; 456b1b3d7a2SStefano Zampini if (!n_added) break; 457b1b3d7a2SStefano Zampini } 458b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 459b1b3d7a2SStefano Zampini 460883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 461a9b99552SStefano 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); 462b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 463a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 464883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 465883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 466b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 467b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 468a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 469b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 470b1b3d7a2SStefano Zampini 471b1b3d7a2SStefano Zampini /* II block */ 472b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 473b1b3d7a2SStefano Zampini } 474b1b3d7a2SStefano Zampini } else { 475b1b3d7a2SStefano Zampini PetscInt n_I; 476b1b3d7a2SStefano Zampini 477b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 478b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 479a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 480b1b3d7a2SStefano Zampini 481b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 482883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 483b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 484b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 485b1b3d7a2SStefano Zampini 486b1b3d7a2SStefano Zampini /* II block is the same */ 487b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 488b1b3d7a2SStefano Zampini AE_II = A_II; 489b1b3d7a2SStefano Zampini } 490b1b3d7a2SStefano Zampini } 4915a95e1ceSStefano Zampini 492883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 4935a95e1ceSStefano Zampini max_subset_size = 0; 494883469d8SStefano Zampini local_size = 0; 4955a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4965a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 4975a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 498883469d8SStefano Zampini local_size += subset_size; 499883469d8SStefano Zampini } 500883469d8SStefano Zampini 501883469d8SStefano Zampini /* Work arrays for local indices */ 502883469d8SStefano Zampini extra = 0; 50381ea8064SStefano Zampini if (sub_schurs->use_mumps && is_I_layer) { 504a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 505883469d8SStefano Zampini } 506883469d8SStefano Zampini ierr = PetscMalloc1(local_size+extra,&all_local_idx_N);CHKERRQ(ierr); 507883469d8SStefano Zampini if (extra) { 508883469d8SStefano Zampini const PetscInt *idxs; 509a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 510883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 511a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 512883469d8SStefano Zampini } 513883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 514dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr); 515dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 516883469d8SStefano Zampini 517883469d8SStefano Zampini /* Get local indices in local numbering */ 518883469d8SStefano Zampini local_size = 0; 5195a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 520883469d8SStefano Zampini PetscInt j; 521883469d8SStefano Zampini const PetscInt *idxs; 522883469d8SStefano Zampini 5235a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5245a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 525eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 526eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 527eb595f79SStefano Zampini auxnum2[i] = subset_size; 528883469d8SStefano Zampini /* subset indices in local numbering */ 529883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 5305a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 531883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 532883469d8SStefano Zampini local_size += subset_size; 533883469d8SStefano Zampini } 534883469d8SStefano Zampini 5355a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 536d2627357SStefano Zampini Bwork = NULL; 537d2627357SStefano Zampini pivots = NULL; 53806a4e24aSStefano Zampini if (local_size && !benign_trick && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 539d2627357SStefano Zampini PetscScalar lwork; 540d2627357SStefano Zampini 54106a4e24aSStefano Zampini use_getr = PETSC_TRUE; 542d2627357SStefano Zampini B_lwork = -1; 543d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 544d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 545d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 546d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 547d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 548d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 549d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 550d2627357SStefano Zampini } 551d2627357SStefano Zampini 552d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 553dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr); 554dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr); 555dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 556dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr); 557dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr); 558dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets);CHKERRQ(ierr); 559dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr); 560dc456d91SStefano Zampini ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr); 561dc456d91SStefano Zampini if (i != local_size) { 562dc456d91SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %d != %d",i,local_size); 563eb595f79SStefano Zampini } 564dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(all_subsets_n,&l2gmap_subsets);CHKERRQ(ierr); 565e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,NULL,&work_mat);CHKERRQ(ierr); 566d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 567d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 568d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 569d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 5702972d61bSStefano Zampini 5715a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 5725a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 5735a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 5745a95e1ceSStefano Zampini 5755a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 5765a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 5775a95e1ceSStefano Zampini if (subset_size != local_size) { 5785a95e1ceSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 5795a95e1ceSStefano Zampini } 5805a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 581b1b3d7a2SStefano Zampini } 582b1b3d7a2SStefano Zampini 5835a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 5845a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 5855a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 5865a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 5875a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 5885a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 589aa83b6aeSStefano Zampini } 590b1b3d7a2SStefano Zampini 5915a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 592be83ff47SStefano Zampini F = NULL; 5935a95e1ceSStefano Zampini if (!sub_schurs->use_mumps) { 5945a95e1ceSStefano Zampini Mat S_Ej_expl; 5955a95e1ceSStefano Zampini PetscScalar *work; 5965a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 5975a95e1ceSStefano Zampini PetscBool Sdense; 5985a95e1ceSStefano Zampini 5995a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 6005a95e1ceSStefano Zampini local_size = 0; 601b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6025a95e1ceSStefano Zampini IS is_subset_B; 6035a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 6045a95e1ceSStefano Zampini 6055a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 6065a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 6075a95e1ceSStefano Zampini /* EE block */ 6085a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 6095a95e1ceSStefano Zampini /* IE block */ 6105a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 6115a95e1ceSStefano Zampini /* EI block */ 6125a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 6135a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 6145a95e1ceSStefano Zampini } else { 6155a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 6165a95e1ceSStefano Zampini } 617a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 6185a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 6195a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 6205a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 6215a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 622b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 623b1b3d7a2SStefano Zampini KSP ksp; 624b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 6255a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 626b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 627b1b3d7a2SStefano Zampini KSP origksp,schurksp; 628b1b3d7a2SStefano Zampini PC origpc,schurpc; 629b1b3d7a2SStefano Zampini KSPType ksp_type; 630b1b3d7a2SStefano Zampini PetscInt n_internal; 6315a95e1ceSStefano Zampini PetscBool ispcnone; 632b1b3d7a2SStefano Zampini 633b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 6345a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 635b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 636b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 637b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 638b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 6395a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 6405a95e1ceSStefano Zampini if (!ispcnone) { 6415a95e1ceSStefano Zampini PCType pc_type; 642b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 643b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 6445a95e1ceSStefano Zampini } else { 6455a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 6465a95e1ceSStefano Zampini } 647b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 648b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 649b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 650b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 651b1b3d7a2SStefano Zampini if (solver) { 652b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 653b1b3d7a2SStefano Zampini } 654b1b3d7a2SStefano Zampini } 655b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 656b1b3d7a2SStefano Zampini } 6575a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6585a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 6595a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 6605a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 6615a95e1ceSStefano Zampini if (Sdense) { 6625a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 6635a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 664b1b3d7a2SStefano Zampini } 6655a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 6665a95e1ceSStefano Zampini } else { 6675a95e1ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 6685a95e1ceSStefano Zampini } 6695a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 670a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 6715a95e1ceSStefano Zampini local_size += subset_size; 6725a95e1ceSStefano Zampini } 6735a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 674b1b3d7a2SStefano Zampini /* free */ 675b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 676b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 6775a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 678883469d8SStefano Zampini } else { 679be83ff47SStefano Zampini Mat A; 680883469d8SStefano Zampini IS is_A_all; 681be83ff47SStefano Zampini PetscScalar *work,*S_data; 682be83ff47SStefano Zampini PetscInt *idxs_schur,n_I,n_I_all,*dummy_idx,size_schur,size_active_schur,cum,cum2; 683*a0b0af32SStefano Zampini PetscBool mumps_S,S_lower_triangular = PETSC_FALSE; 684883469d8SStefano Zampini 685883469d8SStefano Zampini /* get working mat */ 68681ea8064SStefano Zampini n_I = 0; 68781ea8064SStefano Zampini if (is_I_layer) { 688a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 68981ea8064SStefano Zampini } 690d62866d3SStefano Zampini if (!sub_schurs->is_dir) { 691883469d8SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&is_A_all);CHKERRQ(ierr); 692d62866d3SStefano Zampini size_schur = local_size; 693d62866d3SStefano Zampini } else { 694d62866d3SStefano Zampini IS list[2]; 695d62866d3SStefano Zampini 696d62866d3SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size+n_I,all_local_idx_N,PETSC_COPY_VALUES,&list[0]);CHKERRQ(ierr); 697d62866d3SStefano Zampini list[1] = sub_schurs->is_dir; 698d62866d3SStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,2,list,&is_A_all);CHKERRQ(ierr); 699d62866d3SStefano Zampini ierr = ISDestroy(&list[0]);CHKERRQ(ierr); 700d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&size_schur);CHKERRQ(ierr); 701d62866d3SStefano Zampini size_schur += local_size; 702d62866d3SStefano Zampini } 70353892102SStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 70453892102SStefano Zampini size_active_schur = local_size; /* size active schurs does not count any dirichlet dof on the interface */ 7056816873aSStefano Zampini ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 706a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 7073301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr); 7083301b35fSStefano Zampini ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 7093301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 710883469d8SStefano Zampini 711d47842beSStefano Zampini /* when using te benign subspace trick, the local Schur complements are SPD */ 712d47842beSStefano Zampini if (benign_trick) sub_schurs->is_posdef = PETSC_TRUE; 713d47842beSStefano Zampini 71408122e43SStefano Zampini if (n_I) { 7159ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 716883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 717883469d8SStefano Zampini } else { 718883469d8SStefano Zampini ierr = MatGetFactor(A,MATSOLVERMUMPS,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 719883469d8SStefano Zampini } 720883469d8SStefano Zampini /* subsets ordered last */ 721d62866d3SStefano Zampini ierr = PetscMalloc1(size_schur,&idxs_schur);CHKERRQ(ierr); 722d62866d3SStefano Zampini for (i=0;i<size_schur;i++) { 723883469d8SStefano Zampini idxs_schur[i] = n_I+i+1; 724883469d8SStefano Zampini } 7255a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 726d62866d3SStefano Zampini ierr = MatMumpsSetSchurIndices(F,size_schur,idxs_schur);CHKERRQ(ierr); 7275a95e1ceSStefano Zampini #endif 728883469d8SStefano Zampini ierr = PetscFree(idxs_schur);CHKERRQ(ierr); 729883469d8SStefano Zampini 730883469d8SStefano Zampini /* factorization step */ 7319ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 732883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 733be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 734be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 735*a0b0af32SStefano Zampini if (sub_schurs->is_posdef) { 736*a0b0af32SStefano Zampini ierr = MatMumpsSchurComplementSetSym(F,1);CHKERRQ(ierr); 737*a0b0af32SStefano Zampini } else { 738*a0b0af32SStefano Zampini ierr = MatMumpsSchurComplementSetSym(F,2);CHKERRQ(ierr); 739*a0b0af32SStefano Zampini } 740be83ff47SStefano Zampini #endif 741883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 742*a0b0af32SStefano Zampini S_lower_triangular = PETSC_TRUE; 743883469d8SStefano Zampini } else { 744883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 745be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 746be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 747be83ff47SStefano Zampini #endif 748883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 749883469d8SStefano Zampini } 750883469d8SStefano Zampini 751883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 7525a95e1ceSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 753883469d8SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 7545a95e1ceSStefano Zampini #endif 755d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 756d5574798SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I_all);CHKERRQ(ierr); 757be83ff47SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && (n_I == n_I_all)); 7584a6c6b0dSStefano Zampini mumps_S = PETSC_TRUE; 759be83ff47SStefano Zampini } else { /* we can't use MUMPS when size_schur == size_of_the_problem */ 760be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 761be83ff47SStefano Zampini reuse_solvers = PETSC_FALSE; 7624a6c6b0dSStefano Zampini mumps_S = PETSC_FALSE; 763be83ff47SStefano Zampini } 764be83ff47SStefano Zampini 765be83ff47SStefano Zampini if (reuse_solvers) { 766d62866d3SStefano Zampini Mat A_II; 76753892102SStefano Zampini Vec vec1_B; 768d62866d3SStefano Zampini PCBDDCReuseMumps msolv_ctx; 769d5574798SStefano Zampini 770d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 7716816873aSStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 772e28d306cSStefano Zampini } else { 773e28d306cSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_mumps);CHKERRQ(ierr); 774d62866d3SStefano Zampini } 775e28d306cSStefano Zampini msolv_ctx = sub_schurs->reuse_mumps; 776be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 777be83ff47SStefano Zampini ierr = MatGetSize(A_II,&msolv_ctx->n_I,NULL);CHKERRQ(ierr); 778d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 779d5574798SStefano Zampini msolv_ctx->F = F; 780d5574798SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,&msolv_ctx->rhs);CHKERRQ(ierr); 781d62866d3SStefano Zampini 782d62866d3SStefano Zampini /* interior solver */ 783d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->interior_solver);CHKERRQ(ierr); 784d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 785d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 786d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 787d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolve);CHKERRQ(ierr); 788e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolveTranspose);CHKERRQ(ierr); 789d62866d3SStefano Zampini 790d62866d3SStefano Zampini /* correction solver */ 791d62866d3SStefano Zampini ierr = PCCreate(PETSC_COMM_SELF,&msolv_ctx->correction_solver);CHKERRQ(ierr); 792d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,A,A);CHKERRQ(ierr); 793d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 794d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 795d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolve);CHKERRQ(ierr); 796e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolveTranspose);CHKERRQ(ierr); 79753892102SStefano Zampini 79853892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 79953892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 80053892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 80153892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 80253892102SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 80353892102SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 80453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 80553892102SStefano Zampini msolv_ctx->is_R = is_A_all; 806d5574798SStefano Zampini } 80708122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 80853892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 8095db18549SStefano Zampini 810be83ff47SStefano Zampini /* Work arrays */ 811be83ff47SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 8125a94c880SStefano Zampini 8135a94c880SStefano Zampini /* matrices for adaptive selection */ 81412d906b1SStefano Zampini if (compute_Stilda) { 8155a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_tilda_all) { 8165a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 8175a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 8185a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 8195a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 8205a94c880SStefano Zampini } 8215a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_inv_all) { 8225a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 8235a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 8245a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 8255a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 8265a94c880SStefano Zampini } 82712d906b1SStefano Zampini } 828d2627357SStefano Zampini 829be83ff47SStefano Zampini /* S_Ej_all */ 830be83ff47SStefano Zampini cum = cum2 = 0; 831be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 83265d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 83365d8bf0aSStefano Zampini PetscInt j; 83465d8bf0aSStefano Zampini 8355a95e1ceSStefano Zampini /* get S_E */ 836b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 837*a0b0af32SStefano Zampini if (mumps_S && S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */ 838be83ff47SStefano Zampini PetscInt k; 839be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 840be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 841be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 842be83ff47SStefano Zampini work[j*subset_size+k] = S_data[cum2+k*size_schur+j]; 843be83ff47SStefano Zampini } 844be83ff47SStefano Zampini } 84506a4e24aSStefano Zampini } else { /* just copy to workspace */ 846be83ff47SStefano Zampini PetscInt k; 847be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 848be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 849be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 850be83ff47SStefano Zampini } 851be83ff47SStefano Zampini } 8529087bf02SStefano Zampini } 8535a95e1ceSStefano Zampini /* insert S_E values */ 854be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 8555a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 856a1337663SStefano Zampini 857be83ff47SStefano Zampini /* if adaptivity is requested, invert S_E block */ 858862806e4SStefano Zampini if (compute_Stilda) { 85908122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 86008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 86106a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 8625a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 8632972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 8645a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 8652972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 866d6462365SStefano Zampini } else { 867d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 868d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 869d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 870d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 8712972d61bSStefano Zampini } 87208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 8735a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8749087bf02SStefano Zampini } 875be83ff47SStefano Zampini cum += subset_size; 876be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 877be83ff47SStefano Zampini } 878be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 8794a6c6b0dSStefano Zampini 880be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 8814a6c6b0dSStefano Zampini if (mumps_S) { 882be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 8834a6c6b0dSStefano Zampini } 8844a6c6b0dSStefano Zampini #endif 88545951f25SStefano Zampini if (compute_Stilda && size_active_schur) { 8864a6c6b0dSStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur) { /* we already computed the inverse */ 8874a6c6b0dSStefano Zampini PetscInt j; 8884a6c6b0dSStefano Zampini for (j=0;j<size_schur;j++) dummy_idx[j] = j; 8895a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8904a6c6b0dSStefano Zampini } else { 8914a6c6b0dSStefano Zampini if (mumps_S) { /* use MatMumps calls to invert S */ 8924a6c6b0dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 893be83ff47SStefano Zampini ierr = MatMumpsInvertSchurComplement(F);CHKERRQ(ierr); 894be83ff47SStefano Zampini ierr = MatMumpsGetSchurComplement(F,&S_all);CHKERRQ(ierr); 8954a6c6b0dSStefano Zampini #endif 8964a6c6b0dSStefano Zampini } else { /* we need to invert explicitly since we are not using MUMPS for S */ 897be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 898be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 899be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 90006a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 901be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 902be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 903be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 904be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 905d6462365SStefano Zampini } else { 906d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 907d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 908d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 909d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 910be83ff47SStefano Zampini } 911be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 912be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 913be83ff47SStefano Zampini } 914be83ff47SStefano Zampini /* S_Ej_tilda_all */ 915be83ff47SStefano Zampini cum = cum2 = 0; 916be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 917be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 918be83ff47SStefano Zampini PetscInt j; 919862806e4SStefano Zampini 920862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 921be83ff47SStefano Zampini /* get (St^-1)_E */ 92206a4e24aSStefano Zampini /* Here I don't need to expand to upper triangular since St^-1 92306a4e24aSStefano Zampini will be properly accessed later during adaptive selection */ 924*a0b0af32SStefano Zampini if (S_lower_triangular) { 925be83ff47SStefano Zampini PetscInt k; 926be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 927be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 928be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 929be83ff47SStefano Zampini } 930be83ff47SStefano Zampini } 931be83ff47SStefano Zampini } else { /* copy to workspace */ 932be83ff47SStefano Zampini PetscInt k; 933be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 934be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 935be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 936be83ff47SStefano Zampini } 937be83ff47SStefano Zampini } 938be83ff47SStefano Zampini } 939be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 9405a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 941be83ff47SStefano Zampini cum += subset_size; 942be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 943883469d8SStefano Zampini } 944be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 945be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 9464a6c6b0dSStefano Zampini if (mumps_S) { 947be83ff47SStefano Zampini ierr = MatMumpsRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 9485db18549SStefano Zampini } 949be83ff47SStefano Zampini #endif 9509087bf02SStefano Zampini } 9514a6c6b0dSStefano Zampini } 9524a6c6b0dSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 9534a6c6b0dSStefano Zampini } 9545a94c880SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 955be83ff47SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 956a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 957a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 958a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 959a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 960a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 9615db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9625db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9635a95e1ceSStefano Zampini if (compute_Stilda) { 9645a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9655a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9665a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9675a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 96808122e43SStefano Zampini } 969a1337663SStefano Zampini 9705db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 9715db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 9723927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 9733927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 9745a95e1ceSStefano Zampini 9755db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 9765a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_all) { 977dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_INITIAL_MATRIX,&submats);CHKERRQ(ierr); 9785a94c880SStefano Zampini sub_schurs->sum_S_Ej_all = submats[0]; 9795a94c880SStefano Zampini } else { 9805a94c880SStefano Zampini ierr = PetscMalloc1(1,&submats);CHKERRQ(ierr); 9815a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_all; 982dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 9835a94c880SStefano Zampini } 98408122e43SStefano Zampini 985ac632422SStefano Zampini /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */ 986ac632422SStefano Zampini if (faster_deluxe) { 9875a95e1ceSStefano Zampini Mat tmpmat; 9885a95e1ceSStefano Zampini PetscScalar *array; 9895a95e1ceSStefano Zampini PetscInt cum; 9905a95e1ceSStefano Zampini 9915a95e1ceSStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 9925a95e1ceSStefano Zampini cum = 0; 9935a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 9945a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 9955a95e1ceSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 9965a95e1ceSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 99706a4e24aSStefano Zampini if (!use_getr) { 9985a95e1ceSStefano Zampini PetscInt j,k; 9995a95e1ceSStefano Zampini 10005a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 10015a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 10025a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 10035a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 10045a95e1ceSStefano Zampini for (j=0;j<B_N;j++) { 10055a95e1ceSStefano Zampini for (k=j+1;k<B_N;k++) { 10065a95e1ceSStefano Zampini array[k*B_N+j+cum] = array[j*B_N+k+cum]; 10075a95e1ceSStefano Zampini } 10085a95e1ceSStefano Zampini } 1009d6462365SStefano Zampini } else { 1010d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 1011d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1012d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1013d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 10145a95e1ceSStefano Zampini } 10155a95e1ceSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10165a95e1ceSStefano Zampini cum += subset_size*subset_size; 10175a95e1ceSStefano Zampini } 10185a95e1ceSStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 10195a95e1ceSStefano Zampini ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr); 10205a95e1ceSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 1021ac632422SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 10225a95e1ceSStefano Zampini sub_schurs->S_Ej_all = tmpmat; 10235a95e1ceSStefano Zampini } 10245a95e1ceSStefano Zampini 1025f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 10265a95e1ceSStefano Zampini if (compute_Stilda) { 10275a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 1028a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 10295a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_tilda_all; 1030dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 10315a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 103208122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 10335a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_inv_all; 1034dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 103508122e43SStefano Zampini } 10363202ece2SStefano Zampini 10375a95e1ceSStefano Zampini /* free workspace */ 10385a94c880SStefano Zampini ierr = PetscFree(submats);CHKERRQ(ierr); 103906a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 1040a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 10413202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1042dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 10435a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 1044b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1045b1b3d7a2SStefano Zampini } 1046b1b3d7a2SStefano Zampini 1047b1b3d7a2SStefano Zampini #undef __FUNCT__ 1048b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 1049a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 1050b1b3d7a2SStefano Zampini { 10519bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 10525a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1053b1b3d7a2SStefano Zampini PetscBool is_sorted; 1054b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1055b1b3d7a2SStefano Zampini 1056b1b3d7a2SStefano Zampini PetscFunctionBegin; 1057b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 1058b1b3d7a2SStefano Zampini if (!is_sorted) { 1059b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1060b1b3d7a2SStefano Zampini } 1061b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 1062b1b3d7a2SStefano Zampini if (!is_sorted) { 1063b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1064b1b3d7a2SStefano Zampini } 1065b1b3d7a2SStefano Zampini 1066b1b3d7a2SStefano Zampini /* reset any previous data */ 1067b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1068b1b3d7a2SStefano Zampini 10695a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 10709bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1071b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 107208122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 1073b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1074b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 1075b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1076b1b3d7a2SStefano Zampini } 1077b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 1078b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 107908122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1080b1b3d7a2SStefano Zampini } 1081b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 1082b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 1083d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1084d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1085b1b3d7a2SStefano Zampini 10866816873aSStefano Zampini /* Determine if MUMPS can be used */ 1087883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 1088883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1089a64f4aa4SStefano Zampini sub_schurs->use_mumps = PETSC_TRUE; 1090883469d8SStefano Zampini #endif 1091b1b3d7a2SStefano Zampini 1092b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1093b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1094b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1095b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 10965db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 10975db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 10985db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 10995db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 11005a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1101b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1102a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { /* sort by local ordering mumps is not present */ 1103b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1104b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 1105b96c3477SStefano Zampini } 11069bb4a8caSStefano Zampini } 1107d62866d3SStefano Zampini sub_schurs->is_vertices = vertices; 1108b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1109b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 111008122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1111b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1112b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1113b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1114b1b3d7a2SStefano Zampini } 1115b1b3d7a2SStefano Zampini 1116b1b3d7a2SStefano Zampini #undef __FUNCT__ 111734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 111834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 111934a97f8cSStefano Zampini { 112034a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 112134a97f8cSStefano Zampini PetscErrorCode ierr; 112234a97f8cSStefano Zampini 112334a97f8cSStefano Zampini PetscFunctionBegin; 112434a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 11255ff63025SStefano Zampini schurs_ctx->n_subs = 0; 112634a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 112734a97f8cSStefano Zampini PetscFunctionReturn(0); 112834a97f8cSStefano Zampini } 112934a97f8cSStefano Zampini 113034a97f8cSStefano Zampini #undef __FUNCT__ 113134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 113234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 113334a97f8cSStefano Zampini { 113434a97f8cSStefano Zampini PetscErrorCode ierr; 113534a97f8cSStefano Zampini 113634a97f8cSStefano Zampini PetscFunctionBegin; 113734a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 113834a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 113934a97f8cSStefano Zampini PetscFunctionReturn(0); 114034a97f8cSStefano Zampini } 114134a97f8cSStefano Zampini 114234a97f8cSStefano Zampini #undef __FUNCT__ 114334a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 114434a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 114534a97f8cSStefano Zampini { 114634a97f8cSStefano Zampini PetscInt i; 114734a97f8cSStefano Zampini PetscErrorCode ierr; 114834a97f8cSStefano Zampini 114934a97f8cSStefano Zampini PetscFunctionBegin; 11501e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1151b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1152b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1153b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 11545db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 11555db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 115641c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 115741c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 115808122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1159a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 11605db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 1161d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 1162d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 116308122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 116434a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1165b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 116634a97f8cSStefano Zampini } 11675ff63025SStefano Zampini if (sub_schurs->n_subs) { 1168b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 11693dc780c3SStefano Zampini } 1170d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1171d62866d3SStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 1172d62866d3SStefano Zampini } 1173d62866d3SStefano Zampini ierr = PetscFree(sub_schurs->reuse_mumps);CHKERRQ(ierr); 117434a97f8cSStefano Zampini sub_schurs->n_subs = 0; 117534a97f8cSStefano Zampini PetscFunctionReturn(0); 117634a97f8cSStefano Zampini } 117734a97f8cSStefano Zampini 117834a97f8cSStefano Zampini #undef __FUNCT__ 117934a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 11802a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 118134a97f8cSStefano Zampini { 118234a97f8cSStefano Zampini PetscInt i,j,n; 118334a97f8cSStefano Zampini PetscErrorCode ierr; 118434a97f8cSStefano Zampini 118534a97f8cSStefano Zampini PetscFunctionBegin; 118634a97f8cSStefano Zampini n = 0; 118734a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 118834a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 118934a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 119034a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 119134a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 119234a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 119334a97f8cSStefano Zampini queue_tip[n] = dof; 119434a97f8cSStefano Zampini n++; 119534a97f8cSStefano Zampini } 119634a97f8cSStefano Zampini } 119734a97f8cSStefano Zampini } 119834a97f8cSStefano Zampini *n_added = n; 119934a97f8cSStefano Zampini PetscFunctionReturn(0); 120034a97f8cSStefano Zampini } 1201