#include <petscksp.h>
#include <petsc/private/petscimpl.h>

static char help[] = "Solves a linear system using PCHPDDM.\n\n";

int main(int argc, char **args)
{
  Vec             b;            /* computed solution and RHS */
  Mat             A, aux, X, B; /* linear system matrix */
  KSP             ksp;          /* linear solver context */
  PC              pc;
  IS              is, sizes;
  const PetscInt *idx;
  PetscMPIInt     rank, size;
  PetscInt        m, N = 1;
  PetscLayout     map;
  PetscViewer     viewer;
  char            dir[PETSC_MAX_PATH_LEN], name[PETSC_MAX_PATH_LEN], type[256];
  PetscBool3      share = PETSC_BOOL3_UNKNOWN;
  PetscBool       flg, set, transpose = PETSC_FALSE, explicit = PETSC_FALSE;

  PetscFunctionBeginUser;
  PetscCall(PetscInitialize(&argc, &args, NULL, help));
  PetscCall(PetscLogDefaultBegin());
  PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
  PetscCheck(size == 4, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This example requires 4 processes");
  PetscCall(PetscOptionsGetInt(NULL, NULL, "-rhs", &N, NULL));
  PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
  PetscCall(MatCreate(PETSC_COMM_WORLD, &A));
  PetscCall(PetscStrncpy(dir, ".", sizeof(dir)));
  PetscCall(PetscOptionsGetString(NULL, NULL, "-load_dir", dir, sizeof(dir), NULL));
  /* loading matrices */
  PetscCall(PetscSNPrintf(name, sizeof(name), "%s/sizes_%d.dat", dir, size));
  PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
  PetscCall(ISCreate(PETSC_COMM_WORLD, &sizes));
  PetscCall(ISLoad(sizes, viewer));
  PetscCall(ISGetIndices(sizes, &idx));
  PetscCall(MatSetSizes(A, idx[0], idx[1], idx[2], idx[3]));
  PetscCall(MatCreate(PETSC_COMM_WORLD, &X));
  PetscCall(MatSetSizes(X, idx[4], idx[4], PETSC_DETERMINE, PETSC_DETERMINE));
  PetscCall(MatSetUp(X));
  PetscCall(ISRestoreIndices(sizes, &idx));
  PetscCall(ISDestroy(&sizes));
  PetscCall(PetscViewerDestroy(&viewer));
  PetscCall(PetscSNPrintf(name, sizeof(name), "%s/A.dat", dir));
  PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
  PetscCall(MatLoad(A, viewer));
  PetscCall(PetscViewerDestroy(&viewer));
  PetscCall(PetscSNPrintf(name, sizeof(name), "%s/is_%d.dat", dir, size));
  PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
  PetscCall(ISCreate(PETSC_COMM_WORLD, &sizes));
  PetscCall(MatGetLayouts(X, &map, NULL));
  PetscCall(ISSetLayout(sizes, map));
  PetscCall(ISLoad(sizes, viewer));
  PetscCall(ISGetLocalSize(sizes, &m));
  PetscCall(ISGetIndices(sizes, &idx));
  PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_COPY_VALUES, &is));
  PetscCall(ISRestoreIndices(sizes, &idx));
  PetscCall(ISDestroy(&sizes));
  PetscCall(MatGetBlockSize(A, &m));
  PetscCall(ISSetBlockSize(is, m));
  PetscCall(PetscViewerDestroy(&viewer));
  PetscCall(PetscSNPrintf(name, sizeof(name), "%s/Neumann_%d.dat", dir, size));
  PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
  PetscCall(MatLoad(X, viewer));
  PetscCall(PetscViewerDestroy(&viewer));
  PetscCall(MatGetDiagonalBlock(X, &B));
  PetscCall(MatDuplicate(B, MAT_COPY_VALUES, &aux));
  PetscCall(MatDestroy(&X));
  flg = PETSC_FALSE;
  PetscCall(PetscOptionsGetBool(NULL, NULL, "-pc_hpddm_levels_1_st_share_sub_ksp", &flg, &set));
  if (flg) { /* PETSc LU/Cholesky is struggling numerically for bs > 1          */
             /* only set the proper bs for the geneo_share_* tests, 1 otherwise */
    PetscCall(MatSetBlockSizesFromMats(aux, A, A));
    share = PETSC_BOOL3_TRUE;
  } else if (set) share = PETSC_BOOL3_FALSE;
  PetscCall(MatSetOption(A, MAT_SYMMETRIC, PETSC_TRUE));
  PetscCall(MatSetOption(aux, MAT_SYMMETRIC, PETSC_TRUE));
  /* ready for testing */
  PetscOptionsBegin(PETSC_COMM_WORLD, "", "", "");
  PetscCall(PetscStrncpy(type, MATAIJ, sizeof(type)));
  PetscCall(PetscOptionsFList("-mat_type", "Matrix type", "MatSetType", MatList, type, type, 256, &flg));
  PetscOptionsEnd();
  PetscCall(MatConvert(A, type, MAT_INPLACE_MATRIX, &A));
  PetscCall(MatConvert(aux, type, MAT_INPLACE_MATRIX, &aux));
  PetscCall(KSPCreate(PETSC_COMM_WORLD, &ksp));
  PetscCall(KSPSetOperators(ksp, A, A));
  PetscCall(KSPGetPC(ksp, &pc));
  PetscCall(PCSetType(pc, PCHPDDM));
#if defined(PETSC_HAVE_HPDDM) && defined(PETSC_HAVE_DYNAMIC_LIBRARIES) && defined(PETSC_USE_SHARED_LIBRARIES)
  flg = PETSC_FALSE;
  PetscCall(PetscOptionsGetBool(NULL, NULL, "-reset", &flg, NULL));
  if (flg) {
    PetscCall(PetscOptionsSetValue(NULL, "-pc_hpddm_block_splitting", "true"));
    PetscCall(PCSetFromOptions(pc));
    PetscCall(PCSetUp(pc));
    PetscCall(PetscOptionsClearValue(NULL, "-pc_hpddm_block_splitting"));
  }
  PetscCall(PCHPDDMSetAuxiliaryMat(pc, is, aux, NULL, NULL));
  PetscCall(PCHPDDMHasNeumannMat(pc, PETSC_FALSE)); /* PETSC_TRUE is fine as well, just testing */
  if (share == PETSC_BOOL3_UNKNOWN) PetscCall(PCHPDDMSetSTShareSubKSP(pc, PetscBool3ToBool(share)));
  flg = PETSC_FALSE;
  PetscCall(PetscOptionsGetBool(NULL, NULL, "-set_rhs", &flg, NULL));
  if (flg) {          /* user-provided RHS for concurrent generalized eigenvalue problems                          */
    Mat      a, c, P; /* usually assembled automatically in PCHPDDM, this is solely for testing PCHPDDMSetRHSMat() */
    PetscInt rstart, rend, location;

    PetscCall(MatDuplicate(aux, MAT_DO_NOT_COPY_VALUES, &B)); /* duplicate so that MatStructure is SAME_NONZERO_PATTERN */
    PetscCall(MatGetDiagonalBlock(A, &a));
    PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
    PetscCall(ISGetLocalSize(is, &m));
    PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, rend - rstart, m, 1, NULL, &P));
    for (m = rstart; m < rend; ++m) {
      PetscCall(ISLocate(is, m, &location));
      PetscCheck(location >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "IS of the auxiliary Mat does not include all local rows of A");
      PetscCall(MatSetValue(P, m - rstart, location, 1.0, INSERT_VALUES));
    }
    PetscCall(MatAssemblyBegin(P, MAT_FINAL_ASSEMBLY));
    PetscCall(MatAssemblyEnd(P, MAT_FINAL_ASSEMBLY));
    PetscCall(PetscObjectTypeCompare((PetscObject)a, MATSEQAIJ, &flg));
    if (flg) PetscCall(MatPtAP(a, P, MAT_INITIAL_MATRIX, 1.0, &X)); // MatPtAP() is used to extend diagonal blocks with zeros on the overlap
    else {                                                          // workaround for MatPtAP() limitations with some types
      PetscCall(MatConvert(a, MATSEQAIJ, MAT_INITIAL_MATRIX, &c));
      PetscCall(MatPtAP(c, P, MAT_INITIAL_MATRIX, 1.0, &X));
      PetscCall(MatDestroy(&c));
    }
    PetscCall(MatDestroy(&P));
    PetscCall(MatAXPY(B, 1.0, X, SUBSET_NONZERO_PATTERN));
    PetscCall(MatDestroy(&X));
    PetscCall(MatSetOption(B, MAT_SYMMETRIC, PETSC_TRUE));
    PetscCall(PCHPDDMSetRHSMat(pc, B));
    PetscCall(MatDestroy(&B));
  }
#else
  (void)share;
#endif
  PetscCall(MatDestroy(&aux));
  PetscCall(KSPSetFromOptions(ksp));
  PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCASM, &flg));
  if (flg) {
    flg = PETSC_FALSE;
    PetscCall(PetscOptionsGetBool(NULL, NULL, "-pc_hpddm_define_subdomains", &flg, NULL));
    if (flg) {
      IS rows;

      PetscCall(MatGetOwnershipIS(A, &rows, NULL));
      PetscCall(PCASMSetLocalSubdomains(pc, 1, &is, &rows));
      PetscCall(ISDestroy(&rows));
    }
  }
  PetscCall(ISDestroy(&is));
  PetscCall(MatCreateVecs(A, NULL, &b));
  PetscCall(VecSet(b, 1.0));
  PetscCall(PetscOptionsGetBool(NULL, NULL, "-transpose", &transpose, NULL));
  if (!transpose) PetscCall(KSPSolve(ksp, b, b));
  else {
    PetscCall(KSPSolveTranspose(ksp, b, b));
    PetscCall(PetscOptionsGetBool(NULL, NULL, "-ksp_use_explicittranspose", &explicit, NULL));
    if (explicit) PetscCall(KSPSetOperators(ksp, A, A)); /* -ksp_use_explicittranspose does not cache the initial Mat and will transpose the explicit transpose again if not set back to the original Mat */
  }
  PetscCall(VecGetLocalSize(b, &m));
  PetscCall(VecDestroy(&b));
  if (N > 1) {
    KSPType type;
    VecType vt;

    PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
    PetscCall(KSPSetFromOptions(ksp));
    PetscCall(MatGetVecType(A, &vt));
    PetscCall(MatCreateDenseFromVecType(PETSC_COMM_WORLD, vt, m, PETSC_DECIDE, PETSC_DECIDE, N, PETSC_DECIDE, NULL, &B));
    PetscCall(MatCreateDenseFromVecType(PETSC_COMM_WORLD, vt, m, PETSC_DECIDE, PETSC_DECIDE, N, PETSC_DECIDE, NULL, &X));
    PetscCall(MatSetRandom(B, NULL));
    /* this is algorithmically optimal in the sense that blocks of vectors are coarsened or interpolated using matrix--matrix operations */
    /* PCHPDDM however heavily relies on MPI[S]BAIJ format for which there is no efficient MatProduct implementation */
    if (!transpose) PetscCall(KSPMatSolve(ksp, B, X));
    else {
      PetscCall(KSPMatSolveTranspose(ksp, B, X));
      if (explicit) PetscCall(KSPSetOperators(ksp, A, A)); /* same as in the prior KSPSolveTranspose() */
    }
    PetscCall(KSPGetType(ksp, &type));
    PetscCall(PetscStrcmp(type, KSPHPDDM, &flg));
#if defined(PETSC_HAVE_HPDDM)
    if (flg) {
      PetscReal    norm;
      KSPHPDDMType type;

      PetscCall(KSPHPDDMGetType(ksp, &type));
      if (type == KSP_HPDDM_TYPE_PREONLY || type == KSP_HPDDM_TYPE_CG || type == KSP_HPDDM_TYPE_GMRES || type == KSP_HPDDM_TYPE_GCRODR) {
        Mat C;

        PetscCall(MatDuplicate(X, MAT_DO_NOT_COPY_VALUES, &C));
        PetscCall(KSPSetMatSolveBatchSize(ksp, 1));
        if (!transpose) PetscCall(KSPMatSolve(ksp, B, C));
        else {
          PetscCall(KSPMatSolveTranspose(ksp, B, C));
          if (explicit) PetscCall(KSPSetOperators(ksp, A, A)); /* same as in the prior KSPMatSolveTranspose() */
        }
        PetscCall(MatAYPX(C, -1.0, X, SAME_NONZERO_PATTERN));
        PetscCall(MatNorm(C, NORM_INFINITY, &norm));
        PetscCall(MatDestroy(&C));
        PetscCheck(norm <= 100 * PETSC_MACHINE_EPSILON, PetscObjectComm((PetscObject)pc), PETSC_ERR_PLIB, "KSPMatSolve%s() and KSPSolve%s() difference has nonzero norm %g with pseudo-block KSPHPDDMType %s", (transpose ? "Transpose" : ""), (transpose ? "Transpose" : ""), (double)norm, KSPHPDDMTypes[type]);
      }
    }
#endif
    PetscCall(MatDestroy(&X));
    PetscCall(MatDestroy(&B));
  }
  PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCHPDDM, &flg));
#if defined(PETSC_HAVE_HPDDM) && defined(PETSC_HAVE_DYNAMIC_LIBRARIES) && defined(PETSC_USE_SHARED_LIBRARIES)
  if (flg) PetscCall(PCHPDDMGetSTShareSubKSP(pc, &flg));
#endif
  if (flg && PetscDefined(USE_LOG)) {
    PetscCall(PetscOptionsHasName(NULL, NULL, "-pc_hpddm_harmonic_overlap", &flg));
    if (!flg) {
      PetscLogEvent      event;
      PetscEventPerfInfo info1, info2;

      PetscCall(PetscLogEventRegister("MatLUFactorSym", PC_CLASSID, &event));
      PetscCall(PetscLogEventGetPerfInfo(PETSC_DETERMINE, event, &info1));
      PetscCall(PetscLogEventRegister("MatLUFactorNum", PC_CLASSID, &event));
      PetscCall(PetscLogEventGetPerfInfo(PETSC_DETERMINE, event, &info2));
      if (!info1.count && !info2.count) {
        PetscCall(PetscLogEventRegister("MatCholFctrSym", PC_CLASSID, &event));
        PetscCall(PetscLogEventGetPerfInfo(PETSC_DETERMINE, event, &info1));
        PetscCall(PetscLogEventRegister("MatCholFctrNum", PC_CLASSID, &event));
        PetscCall(PetscLogEventGetPerfInfo(PETSC_DETERMINE, event, &info2));
        PetscCheck(info2.count > info1.count, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cholesky numerical factorization (%d) not called more times than Cholesky symbolic factorization (%d), broken -pc_hpddm_levels_1_st_share_sub_ksp", info2.count, info1.count);
      } else PetscCheck(info2.count > info1.count, PETSC_COMM_SELF, PETSC_ERR_PLIB, "LU numerical factorization (%d) not called more times than LU symbolic factorization (%d), broken -pc_hpddm_levels_1_st_share_sub_ksp", info2.count, info1.count);
    }
  }
#if defined(PETSC_HAVE_HPDDM) && defined(PETSC_HAVE_DYNAMIC_LIBRARIES) && defined(PETSC_USE_SHARED_LIBRARIES)
  if (N == 1) {
    flg = PETSC_FALSE;
    PetscCall(PetscOptionsGetBool(NULL, NULL, "-successive_solves", &flg, NULL));
    if (flg) {
      KSPConvergedReason reason[2];
      PetscInt           iterations[3];

      PetscCall(KSPGetConvergedReason(ksp, reason));
      PetscCall(KSPGetTotalIterations(ksp, iterations));
      PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
      PetscCall(KSPSetFromOptions(ksp));
      flg = PETSC_FALSE;
      PetscCall(PetscOptionsGetBool(NULL, NULL, "-pc_hpddm_block_splitting", &flg, NULL));
      if (!flg) {
        PetscCall(PetscSNPrintf(name, sizeof(name), "%s/sizes_%d.dat", dir, size));
        PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
        PetscCall(ISCreate(PETSC_COMM_WORLD, &sizes));
        PetscCall(ISLoad(sizes, viewer));
        PetscCall(ISGetIndices(sizes, &idx));
        PetscCall(MatCreate(PETSC_COMM_WORLD, &X));
        PetscCall(MatSetSizes(X, idx[4], idx[4], PETSC_DETERMINE, PETSC_DETERMINE));
        PetscCall(MatSetUp(X));
        PetscCall(ISRestoreIndices(sizes, &idx));
        PetscCall(ISDestroy(&sizes));
        PetscCall(PetscViewerDestroy(&viewer));
        PetscCall(PetscSNPrintf(name, sizeof(name), "%s/is_%d.dat", dir, size));
        PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
        PetscCall(ISCreate(PETSC_COMM_WORLD, &sizes));
        PetscCall(MatGetLayouts(X, &map, NULL));
        PetscCall(ISSetLayout(sizes, map));
        PetscCall(ISLoad(sizes, viewer));
        PetscCall(ISGetLocalSize(sizes, &m));
        PetscCall(ISGetIndices(sizes, &idx));
        PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_COPY_VALUES, &is));
        PetscCall(ISRestoreIndices(sizes, &idx));
        PetscCall(ISDestroy(&sizes));
        PetscCall(MatGetBlockSize(A, &m));
        PetscCall(ISSetBlockSize(is, m));
        PetscCall(PetscViewerDestroy(&viewer));
        PetscCall(PetscSNPrintf(name, sizeof(name), "%s/Neumann_%d.dat", dir, size));
        PetscCall(PetscViewerBinaryOpen(PETSC_COMM_WORLD, name, FILE_MODE_READ, &viewer));
        PetscCall(MatLoad(X, viewer));
        PetscCall(PetscViewerDestroy(&viewer));
        PetscCall(MatGetDiagonalBlock(X, &B));
        PetscCall(MatDuplicate(B, MAT_COPY_VALUES, &aux));
        PetscCall(MatDestroy(&X));
        PetscCall(MatSetBlockSizesFromMats(aux, A, A));
        PetscCall(MatSetOption(aux, MAT_SYMMETRIC, PETSC_TRUE));
        PetscCall(MatConvert(aux, type, MAT_INPLACE_MATRIX, &aux));
      }
      PetscCall(MatCreateVecs(A, NULL, &b));
      PetscCall(PetscObjectStateIncrease((PetscObject)A));
      if (!flg) PetscCall(PCHPDDMSetAuxiliaryMat(pc, NULL, aux, NULL, NULL));
      PetscCall(VecSet(b, 1.0));
      if (!transpose) PetscCall(KSPSolve(ksp, b, b));
      else PetscCall(KSPSolveTranspose(ksp, b, b));
      PetscCall(KSPGetConvergedReason(ksp, reason + 1));
      PetscCall(KSPGetTotalIterations(ksp, iterations + 1));
      iterations[1] -= iterations[0];
      PetscCheck(reason[0] == reason[1] && PetscAbs(iterations[0] - iterations[1]) <= 3, PetscObjectComm((PetscObject)ksp), PETSC_ERR_PLIB, "Successive calls to KSPSolve%s() did not converge for the same reason (%s v. %s) or with the same number of iterations (+/- 3, %" PetscInt_FMT " v. %" PetscInt_FMT ")", (transpose ? "Transpose" : ""), KSPConvergedReasons[reason[0]], KSPConvergedReasons[reason[1]], iterations[0], iterations[1]);
      PetscCall(PetscObjectStateIncrease((PetscObject)A));
      if (!flg) PetscCall(PCHPDDMSetAuxiliaryMat(pc, is, aux, NULL, NULL));
      PetscCall(PCSetFromOptions(pc));
      PetscCall(VecSet(b, 1.0));
      if (!transpose) PetscCall(KSPSolve(ksp, b, b));
      else PetscCall(KSPSolveTranspose(ksp, b, b));
      PetscCall(KSPGetConvergedReason(ksp, reason + 1));
      PetscCall(KSPGetTotalIterations(ksp, iterations + 2));
      iterations[2] -= iterations[0] + iterations[1];
      PetscCheck(reason[0] == reason[1] && PetscAbs(iterations[0] - iterations[2]) <= 3, PetscObjectComm((PetscObject)ksp), PETSC_ERR_PLIB, "Successive calls to KSPSolve%s() did not converge for the same reason (%s v. %s) or with the same number of iterations (+/- 3, %" PetscInt_FMT " v. %" PetscInt_FMT ")", (transpose ? "Transpose" : ""), KSPConvergedReasons[reason[0]], KSPConvergedReasons[reason[1]], iterations[0], iterations[2]);
      PetscCall(VecDestroy(&b));
      PetscCall(ISDestroy(&is));
      PetscCall(MatDestroy(&aux));
    }
  }
  PetscCall(PetscOptionsGetBool(NULL, NULL, "-viewer", &flg, NULL));
  if (flg) {
    PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCHPDDM, &flg));
    if (flg) {
      PetscCall(PetscStrncpy(dir, "XXXXXX", sizeof(dir)));
      if (rank == 0) PetscCall(PetscMkdtemp(dir));
      PetscCallMPI(MPI_Bcast(dir, 6, MPI_CHAR, 0, PETSC_COMM_WORLD));
      for (PetscInt i = 0; i < 2; ++i) {
        PetscCall(PetscSNPrintf(name, sizeof(name), "%s/%s", dir, i == 0 ? "A" : "A.dat"));
        PetscCall(PetscViewerASCIIOpen(PETSC_COMM_WORLD, name, &viewer));
        PetscCall(PetscViewerPushFormat(viewer, PETSC_VIEWER_ASCII_INFO_DETAIL));
        PetscCall(PCView(pc, viewer));
        PetscCall(PetscViewerPopFormat(viewer));
        PetscCall(PetscViewerDestroy(&viewer));
      }
      PetscCallMPI(MPI_Barrier(PETSC_COMM_WORLD));
      if (rank == 0) PetscCall(PetscRMTree(dir));
    }
  }
#endif
  PetscCall(KSPDestroy(&ksp));
  PetscCall(MatDestroy(&A));
  PetscCall(PetscFinalize());
  return 0;
}

/*TEST

   test:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      nsize: 4
      args: -ksp_rtol 1e-3 -ksp_converged_reason -pc_type {{bjacobi hpddm}shared output} -pc_hpddm_coarse_sub_pc_type lu -sub_pc_type lu -options_left no -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO

   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      suffix: define_subdomains
      nsize: 4
      args: -ksp_rtol 1e-3 -ksp_converged_reason -pc_hpddm_define_subdomains -options_left no -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO
      test:
        args: -pc_type {{asm hpddm}shared output} -pc_hpddm_coarse_sub_pc_type lu -sub_pc_type lu -viewer
      test:
        args: -pc_type hpddm -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_nev 5 -pc_hpddm_coarse_sub_pc_type lu -pc_hpddm_levels_1_sub_pc_type lu -pc_hpddm_coarse_correction none

   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_coarse_pc_type redundant -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO
      test:
        suffix: geneo
        args: -pc_hpddm_coarse_p {{1 2}shared output} -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_nev {{5 15}separate output} -mat_type {{aij baij sbaij}shared output}
      test:
        suffix: geneo_block_splitting
        output_file: output/ex76_geneo_pc_hpddm_levels_1_eps_nev-15.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[6-9]/Linear solve converged due to CONVERGED_RTOL iterations 11/g"
        args: -pc_hpddm_coarse_p 2 -pc_hpddm_levels_1_eps_nev 15 -pc_hpddm_block_splitting -pc_hpddm_levels_1_st_pc_type lu -pc_hpddm_levels_1_eps_gen_non_hermitian -mat_type {{aij baij}shared output} -successive_solves
      test:
        suffix: geneo_share
        output_file: output/ex76_geneo_pc_hpddm_levels_1_eps_nev-5.out
        args: -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_nev 5 -pc_hpddm_levels_1_st_share_sub_ksp -reset {{false true}shared output}
      test:
        suffix: harmonic_overlap_1_define_false
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_pc_type lu -pc_hpddm_levels_1_eps_pc_type lu -pc_hpddm_define_subdomains false -pc_hpddm_levels_1_pc_type asm -pc_hpddm_levels_1_pc_asm_overlap 2 -mat_type baij
      test:
        suffix: harmonic_overlap_1
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_pc_type lu -pc_hpddm_levels_1_eps_pc_type lu -mat_type baij
      test:
        suffix: harmonic_overlap_1_share_petsc
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_pc_type lu -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type petsc -pc_hpddm_levels_1_eps_pc_type lu -mat_type baij
      test:
        requires: mumps
        suffix: harmonic_overlap_1_share_mumps
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type mumps -pc_hpddm_coarse_pc_type cholesky -pc_hpddm_coarse_pc_factor_mat_solver_type mumps -pc_hpddm_coarse_mat_mumps_icntl_15 1
      test:
        requires: mumps
        suffix: harmonic_overlap_1_share_mumps_not_set_explicitly
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_eps_mat_type baij
      test:
        requires: mkl_pardiso
        suffix: harmonic_overlap_1_share_mkl_pardiso
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations [12][0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_eps_mat_type shell -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type mkl_pardiso
      test:
        requires: mkl_pardiso !mumps
        suffix: harmonic_overlap_1_share_mkl_pardiso_no_set_explicitly
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations [12][0-3]/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 1 -pc_hpddm_levels_1_eps_nev 30 -pc_hpddm_levels_1_eps_threshold_relative 1e+1 -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_eps_mat_type shell
      test:
        suffix: harmonic_overlap_2_threshold_relative
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 9/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 2 -pc_hpddm_levels_1_svd_nsv 15 -pc_hpddm_levels_1_svd_threshold_relative 1e-1 -pc_hpddm_levels_1_st_share_sub_ksp -mat_type sbaij
      test:
        suffix: harmonic_overlap_2
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 9/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_harmonic_overlap 2 -pc_hpddm_levels_1_svd_nsv 12 -pc_hpddm_levels_1_svd_type {{trlanczos randomized}shared output} -pc_hpddm_levels_1_st_share_sub_ksp -mat_type sbaij

   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      nsize: 4
      args: -ksp_converged_reason -ksp_max_it 150 -pc_type hpddm -pc_hpddm_levels_1_eps_nev 5 -pc_hpddm_coarse_p 1 -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO -pc_hpddm_define_subdomains
      test:
        suffix: geneo_share_cholesky
        output_file: output/ex76_geneo_share.out
        # extra -pc_hpddm_levels_1_eps_gen_non_hermitian needed to avoid failures with PETSc Cholesky
        args: -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_st_pc_type cholesky -mat_type {{aij sbaij}shared output} -pc_hpddm_levels_1_eps_gen_non_hermitian -pc_hpddm_has_neumann -pc_hpddm_levels_1_st_share_sub_ksp {{false true}shared output} -successive_solves
      test:
        suffix: geneo_share_cholesky_matstructure
        output_file: output/ex76_geneo_share.out
        # extra -pc_hpddm_levels_1_eps_gen_non_hermitian needed to avoid failures with PETSc Cholesky
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 14/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_levels_1_sub_pc_type cholesky -mat_type {{baij sbaij}shared output} -pc_hpddm_levels_1_eps_gen_non_hermitian -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_st_matstructure same -set_rhs {{false true} shared output}
      test:
        suffix: geneo_transpose
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[234]/Linear solve converged due to CONVERGED_RTOL iterations 15/g" -e "s/Linear solve converged due to CONVERGED_RTOL iterations 26/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_gen_non_hermitian -pc_hpddm_has_neumann -pc_hpddm_levels_1_st_share_sub_ksp -successive_solves -transpose -pc_hpddm_coarse_correction {{additive deflated balanced}shared output}
      test:
        suffix: geneo_explicittranspose
        output_file: output/ex76_geneo_share.out
        filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1[234]/Linear solve converged due to CONVERGED_RTOL iterations 15/g" -e "s/Linear solve converged due to CONVERGED_RTOL iterations 26/Linear solve converged due to CONVERGED_RTOL iterations 15/g"
        args: -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_gen_non_hermitian -pc_hpddm_has_neumann -pc_hpddm_levels_1_st_share_sub_ksp -transpose -ksp_use_explicittranspose -rhs 2
      test:
        requires: mumps
        suffix: geneo_share_lu
        output_file: output/ex76_geneo_share.out
        # extra -pc_factor_mat_solver_type mumps needed to avoid failures with PETSc LU
        args: -pc_hpddm_levels_1_sub_pc_type lu -pc_hpddm_levels_1_st_pc_type lu -mat_type baij -pc_hpddm_levels_1_st_pc_factor_mat_solver_type mumps -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type mumps -pc_hpddm_has_neumann -pc_hpddm_levels_1_st_share_sub_ksp {{false true}shared output} -pc_hpddm_coarse_pc_factor_mat_solver_type mumps -pc_hpddm_coarse_mat_mumps_icntl_15 1
      test:
        requires: mumps
        suffix: geneo_share_lu_matstructure
        output_file: output/ex76_geneo_share.out
        # extra -pc_factor_mat_solver_type mumps needed to avoid failures with PETSc LU
        args: -pc_hpddm_levels_1_sub_pc_type lu -mat_type aij -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type mumps -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_levels_1_st_matstructure {{same different}shared output} -pc_hpddm_levels_1_st_pc_type lu -pc_hpddm_levels_1_st_pc_factor_mat_solver_type mumps -successive_solves -pc_hpddm_levels_1_eps_target 1e-5
      test:
        suffix: geneo_share_not_asm
        output_file: output/ex76_geneo_pc_hpddm_levels_1_eps_nev-5.out
        # extra -pc_hpddm_levels_1_eps_gen_non_hermitian needed to avoid failures with PETSc Cholesky
        args: -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_levels_1_eps_gen_non_hermitian -pc_hpddm_has_neumann -pc_hpddm_levels_1_st_share_sub_ksp true -pc_hpddm_levels_1_pc_type gasm -successive_solves

   test:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      suffix: fgmres_geneo_20_p_2
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type lu -pc_hpddm_levels_1_eps_nev 20 -pc_hpddm_coarse_p 2 -pc_hpddm_coarse_pc_type redundant -ksp_type fgmres -pc_hpddm_coarse_mat_type {{baij sbaij}shared output} -pc_hpddm_log_separate {{false true}shared output} -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO

   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      output_file: output/ex76_fgmres_geneo_20_p_2.out
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_eps_nev 20 -pc_hpddm_levels_2_p 2 -pc_hpddm_levels_2_mat_type {{baij sbaij}shared output} -pc_hpddm_levels_2_eps_nev {{5 20}shared output} -pc_hpddm_levels_2_sub_pc_type cholesky -pc_hpddm_levels_2_ksp_type gmres -ksp_type fgmres -pc_hpddm_coarse_mat_type {{baij sbaij}shared output} -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO
      test:
        suffix: fgmres_geneo_20_p_2_geneo
        args: -mat_type {{aij sbaij}shared output}
      test:
        suffix: fgmres_geneo_20_p_2_geneo_algebraic
        args: -pc_hpddm_levels_2_st_pc_type mat
   # PCHPDDM + KSPHPDDM test to exercise multilevel + multiple RHS in one go
   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      output_file: output/ex76_fgmres_geneo_20_p_2.out
      # for -pc_hpddm_coarse_correction additive
      filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 37/Linear solve converged due to CONVERGED_RTOL iterations 25/g"
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_eps_nev 20 -ksp_type hpddm -ksp_hpddm_variant flexible -pc_hpddm_coarse_mat_type baij -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO -rhs 4
      test:
        suffix: fgmres_geneo_20_p_2_geneo_rhs
        args: -pc_hpddm_levels_2_p 2 -pc_hpddm_levels_2_mat_type baij -pc_hpddm_levels_2_eps_nev 5 -pc_hpddm_levels_2_sub_pc_type cholesky -pc_hpddm_levels_2_ksp_max_it 10 -pc_hpddm_levels_2_ksp_type hpddm -pc_hpddm_levels_2_ksp_hpddm_type gmres -mat_type aij -pc_hpddm_coarse_correction {{additive deflated balanced}shared output}
      test:
        requires: cuda
        suffix: fgmres_geneo_20_rhs_cuda
        args: -mat_type aijcusparse -pc_hpddm_coarse_correction {{deflated balanced}shared output}

   testset:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES) mumps defined(PETSC_HAVE_OPENMP_SUPPORT)
      filter: grep -E -e "Linear solve" -e "      executing" | sed -e "s/MPI =      1/MPI =      2/g" -e "s/OMP =      1/OMP =      2/g"
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_eps_nev 15 -pc_hpddm_levels_1_st_pc_type cholesky -pc_hpddm_coarse_p {{1 2}shared output} -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO -pc_hpddm_coarse_pc_factor_mat_solver_type mumps -pc_hpddm_coarse_mat_mumps_icntl_4 2 -pc_hpddm_coarse_mat_mumps_use_omp_threads {{1 2}shared output}
      test:
        suffix: geneo_mumps_use_omp_threads_1
        output_file: output/ex76_geneo_mumps_use_omp_threads.out
        args: -pc_hpddm_coarse_mat_type {{baij sbaij}shared output}
      test:
        suffix: geneo_mumps_use_omp_threads_2
        output_file: output/ex76_geneo_mumps_use_omp_threads.out
        args: -pc_hpddm_coarse_mat_type aij -pc_hpddm_levels_1_eps_threshold_absolute 0.4 -pc_hpddm_coarse_pc_type cholesky -pc_hpddm_coarse_mat_filter 1e-12

   testset: # converge really poorly because of a tiny -pc_hpddm_levels_1_eps_threshold_absolute, but needed for proper code coverage where some subdomains don't call EPSSolve()
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      nsize: 4
      args: -ksp_converged_reason -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_eps_threshold_absolute 0.005 -pc_hpddm_levels_1_eps_use_inertia -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO -pc_hpddm_levels_1_st_share_sub_ksp -pc_hpddm_define_subdomains -pc_hpddm_has_neumann -ksp_rtol 0.9
      filter: sed -e "s/Linear solve converged due to CONVERGED_RTOL iterations 1/Linear solve converged due to CONVERGED_RTOL iterations 141/g"
      test:
        suffix: inertia_petsc
        output_file: output/ex76_1.out
        args: -pc_hpddm_levels_1_sub_pc_factor_mat_solver_type petsc
      test:
        suffix: inertia_mumps
        output_file: output/ex76_1.out
        requires: mumps

   test:
      requires: hpddm slepc datafilespath double !complex !defined(PETSC_USE_64BIT_INDICES) defined(PETSC_HAVE_DYNAMIC_LIBRARIES) defined(PETSC_USE_SHARED_LIBRARIES)
      suffix: reuse_symbolic
      output_file: output/empty.out
      nsize: 4
      args: -pc_type hpddm -pc_hpddm_levels_1_sub_pc_type cholesky -pc_hpddm_levels_1_eps_nev 20 -rhs 4 -pc_hpddm_coarse_correction {{additive deflated balanced}shared output} -ksp_pc_side {{left right}shared output} -ksp_max_it 20 -ksp_type hpddm -load_dir ${DATAFILESPATH}/matrices/hpddm/GENEO -pc_hpddm_define_subdomains -ksp_error_if_not_converged -transpose {{true false} shared output}

TEST*/