seq/mkl_pardiso/mkl_pardiso.c

#include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/
#include <../src/mat/impls/sbaij/seq/sbaij.h>
#include <../src/mat/impls/dense/seq/dense.h>

#if defined(PETSC_HAVE_MKL_INTEL_ILP64)
#define MKL_ILP64
#endif
#include <mkl_pardiso.h>

PETSC_EXTERN void PetscSetMKL_PARDISOThreads(int);

/*
 *  Possible mkl_pardiso phases that controls the execution of the solver.
 *  For more information check mkl_pardiso manual.
 */
#define JOB_ANALYSIS                                                    11
#define JOB_ANALYSIS_NUMERICAL_FACTORIZATION                            12
#define JOB_ANALYSIS_NUMERICAL_FACTORIZATION_SOLVE_ITERATIVE_REFINEMENT 13
#define JOB_NUMERICAL_FACTORIZATION                                     22
#define JOB_NUMERICAL_FACTORIZATION_SOLVE_ITERATIVE_REFINEMENT          23
#define JOB_SOLVE_ITERATIVE_REFINEMENT                                  33
#define JOB_SOLVE_FORWARD_SUBSTITUTION                                  331
#define JOB_SOLVE_DIAGONAL_SUBSTITUTION                                 332
#define JOB_SOLVE_BACKWARD_SUBSTITUTION                                 333
#define JOB_RELEASE_OF_LU_MEMORY                                        0
#define JOB_RELEASE_OF_ALL_MEMORY                                       -1

#define IPARM_SIZE 64

#if defined(PETSC_USE_64BIT_INDICES)
#if defined(PETSC_HAVE_MKL_INTEL_ILP64)
#define INT_TYPE         long long int
#define MKL_PARDISO      pardiso
#define MKL_PARDISO_INIT pardisoinit
#else
/* this is the case where the MKL BLAS/LAPACK are 32 bit integers but the 64 bit integer version of
     of Pardiso code is used; hence the need for the 64 below*/
#define INT_TYPE         long long int
#define MKL_PARDISO      pardiso_64
#define MKL_PARDISO_INIT pardiso_64init
void pardiso_64init(void *pt, INT_TYPE *mtype, INT_TYPE iparm[]) {
  int iparm_copy[IPARM_SIZE], mtype_copy, i;

  mtype_copy = *mtype;
  pardisoinit(pt, &mtype_copy, iparm_copy);
  for (i = 0; i < IPARM_SIZE; i++) iparm[i] = iparm_copy[i];
}
#endif
#else
#define INT_TYPE         int
#define MKL_PARDISO      pardiso
#define MKL_PARDISO_INIT pardisoinit
#endif

/*
 *  Internal data structure.
 *  For more information check mkl_pardiso manual.
 */
typedef struct {
  /* Configuration vector*/
  INT_TYPE iparm[IPARM_SIZE];

  /*
   * Internal mkl_pardiso memory location.
   * After the first call to mkl_pardiso do not modify pt, as that could cause a serious memory leak.
   */
  void *pt[IPARM_SIZE];

  /* Basic mkl_pardiso info*/
  INT_TYPE phase, maxfct, mnum, mtype, n, nrhs, msglvl, err;

  /* Matrix structure*/
  void     *a;
  INT_TYPE *ia, *ja;

  /* Number of non-zero elements*/
  INT_TYPE nz;

  /* Row permutaton vector*/
  INT_TYPE *perm;

  /* Define if matrix preserves sparse structure.*/
  MatStructure matstruc;

  PetscBool needsym;
  PetscBool freeaij;

  /* Schur complement */
  PetscScalar *schur;
  PetscInt     schur_size;
  PetscInt    *schur_idxs;
  PetscScalar *schur_work;
  PetscBLASInt schur_work_size;
  PetscBool    solve_interior;

  /* True if mkl_pardiso function have been used.*/
  PetscBool CleanUp;

  /* Conversion to a format suitable for MKL */
  PetscErrorCode (*Convert)(Mat, PetscBool, MatReuse, PetscBool *, INT_TYPE *, INT_TYPE **, INT_TYPE **, PetscScalar **);
} Mat_MKL_PARDISO;

PetscErrorCode MatMKLPardiso_Convert_seqsbaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v) {
  Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ *)A->data;
  PetscInt      bs = A->rmap->bs, i;

  PetscFunctionBegin;
  PetscCheck(sym, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "This should not happen");
  *v = aa->a;
  if (bs == 1) { /* already in the correct format */
    /* though PetscInt and INT_TYPE are of the same size since they are defined differently the Intel compiler requires a cast */
    *r    = (INT_TYPE *)aa->i;
    *c    = (INT_TYPE *)aa->j;
    *nnz  = (INT_TYPE)aa->nz;
    *free = PETSC_FALSE;
  } else if (reuse == MAT_INITIAL_MATRIX) {
    PetscInt  m = A->rmap->n, nz = aa->nz;
    PetscInt *row, *col;
    PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
    for (i = 0; i < m + 1; i++) row[i] = aa->i[i] + 1;
    for (i = 0; i < nz; i++) col[i] = aa->j[i] + 1;
    *r    = (INT_TYPE *)row;
    *c    = (INT_TYPE *)col;
    *nnz  = (INT_TYPE)nz;
    *free = PETSC_TRUE;
  }
  PetscFunctionReturn(0);
}

PetscErrorCode MatMKLPardiso_Convert_seqbaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v) {
  Mat_SeqBAIJ *aa = (Mat_SeqBAIJ *)A->data;
  PetscInt     bs = A->rmap->bs, i;

  PetscFunctionBegin;
  if (!sym) {
    *v = aa->a;
    if (bs == 1) { /* already in the correct format */
      /* though PetscInt and INT_TYPE are of the same size since they are defined differently the Intel compiler requires a cast */
      *r    = (INT_TYPE *)aa->i;
      *c    = (INT_TYPE *)aa->j;
      *nnz  = (INT_TYPE)aa->nz;
      *free = PETSC_FALSE;
      PetscFunctionReturn(0);
    } else if (reuse == MAT_INITIAL_MATRIX) {
      PetscInt  m = A->rmap->n, nz = aa->nz;
      PetscInt *row, *col;
      PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
      for (i = 0; i < m + 1; i++) row[i] = aa->i[i] + 1;
      for (i = 0; i < nz; i++) col[i] = aa->j[i] + 1;
      *r   = (INT_TYPE *)row;
      *c   = (INT_TYPE *)col;
      *nnz = (INT_TYPE)nz;
    }
    *free = PETSC_TRUE;
  } else {
    SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "This should not happen");
  }
  PetscFunctionReturn(0);
}

PetscErrorCode MatMKLPardiso_Convert_seqaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v) {
  Mat_SeqAIJ  *aa = (Mat_SeqAIJ *)A->data;
  PetscScalar *aav;

  PetscFunctionBegin;
  PetscCall(MatSeqAIJGetArrayRead(A, (const PetscScalar **)&aav));
  if (!sym) { /* already in the correct format */
    *v    = aav;
    *r    = (INT_TYPE *)aa->i;
    *c    = (INT_TYPE *)aa->j;
    *nnz  = (INT_TYPE)aa->nz;
    *free = PETSC_FALSE;
  } else if (reuse == MAT_INITIAL_MATRIX) { /* need to get the triangular part */
    PetscScalar *vals, *vv;
    PetscInt    *row, *col, *jj;
    PetscInt     m = A->rmap->n, nz, i;

    nz = 0;
    for (i = 0; i < m; i++) nz += aa->i[i + 1] - aa->diag[i];
    PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
    PetscCall(PetscMalloc1(nz, &vals));
    jj = col;
    vv = vals;

    row[0] = 0;
    for (i = 0; i < m; i++) {
      PetscInt    *aj = aa->j + aa->diag[i];
      PetscScalar *av = aav + aa->diag[i];
      PetscInt     rl = aa->i[i + 1] - aa->diag[i], j;

      for (j = 0; j < rl; j++) {
        *jj = *aj;
        jj++;
        aj++;
        *vv = *av;
        vv++;
        av++;
      }
      row[i + 1] = row[i] + rl;
    }
    *v    = vals;
    *r    = (INT_TYPE *)row;
    *c    = (INT_TYPE *)col;
    *nnz  = (INT_TYPE)nz;
    *free = PETSC_TRUE;
  } else {
    PetscScalar *vv;
    PetscInt     m = A->rmap->n, i;

    vv = *v;
    for (i = 0; i < m; i++) {
      PetscScalar *av = aav + aa->diag[i];
      PetscInt     rl = aa->i[i + 1] - aa->diag[i], j;
      for (j = 0; j < rl; j++) {
        *vv = *av;
        vv++;
        av++;
      }
    }
    *free = PETSC_TRUE;
  }
  PetscCall(MatSeqAIJRestoreArrayRead(A, (const PetscScalar **)&aav));
  PetscFunctionReturn(0);
}

static PetscErrorCode MatMKLPardisoSolveSchur_Private(Mat F, PetscScalar *B, PetscScalar *X) {
  Mat_MKL_PARDISO     *mpardiso = (Mat_MKL_PARDISO *)F->data;
  Mat                  S, Xmat, Bmat;
  MatFactorSchurStatus schurstatus;

  PetscFunctionBegin;
  PetscCall(MatFactorGetSchurComplement(F, &S, &schurstatus));
  PetscCheck(X != B || schurstatus != MAT_FACTOR_SCHUR_INVERTED, PETSC_COMM_SELF, PETSC_ERR_SUP, "X and B cannot point to the same address");
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mpardiso->schur_size, mpardiso->nrhs, B, &Bmat));
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mpardiso->schur_size, mpardiso->nrhs, X, &Xmat));
  PetscCall(MatSetType(Bmat, ((PetscObject)S)->type_name));
  PetscCall(MatSetType(Xmat, ((PetscObject)S)->type_name));
#if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
  PetscCall(MatBindToCPU(Xmat, S->boundtocpu));
  PetscCall(MatBindToCPU(Bmat, S->boundtocpu));
#endif

#if defined(PETSC_USE_COMPLEX)
  PetscCheck(mpardiso->iparm[12 - 1] != 1, PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "Hermitian solve not implemented yet");
#endif

  switch (schurstatus) {
  case MAT_FACTOR_SCHUR_FACTORED:
    if (!mpardiso->iparm[12 - 1]) {
      PetscCall(MatMatSolve(S, Bmat, Xmat));
    } else { /* transpose solve */
      PetscCall(MatMatSolveTranspose(S, Bmat, Xmat));
    }
    break;
  case MAT_FACTOR_SCHUR_INVERTED:
    PetscCall(MatProductCreateWithMat(S, Bmat, NULL, Xmat));
    if (!mpardiso->iparm[12 - 1]) {
      PetscCall(MatProductSetType(Xmat, MATPRODUCT_AB));
    } else { /* transpose solve */
      PetscCall(MatProductSetType(Xmat, MATPRODUCT_AtB));
    }
    PetscCall(MatProductSetFromOptions(Xmat));
    PetscCall(MatProductSymbolic(Xmat));
    PetscCall(MatProductNumeric(Xmat));
    PetscCall(MatProductClear(Xmat));
    break;
  default: SETERRQ(PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "Unhandled MatFactorSchurStatus %" PetscInt_FMT, F->schur_status); break;
  }
  PetscCall(MatFactorRestoreSchurComplement(F, &S, schurstatus));
  PetscCall(MatDestroy(&Bmat));
  PetscCall(MatDestroy(&Xmat));
  PetscFunctionReturn(0);
}

PetscErrorCode MatFactorSetSchurIS_MKL_PARDISO(Mat F, IS is) {
  Mat_MKL_PARDISO   *mpardiso = (Mat_MKL_PARDISO *)F->data;
  const PetscScalar *arr;
  const PetscInt    *idxs;
  PetscInt           size, i;
  PetscMPIInt        csize;
  PetscBool          sorted;

  PetscFunctionBegin;
  PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)F), &csize));
  PetscCheck(csize <= 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "MKL_PARDISO parallel Schur complements not yet supported from PETSc");
  PetscCall(ISSorted(is, &sorted));
  PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS for MKL_PARDISO Schur complements needs to be sorted");
  PetscCall(ISGetLocalSize(is, &size));
  PetscCall(PetscFree(mpardiso->schur_work));
  PetscCall(PetscBLASIntCast(PetscMax(mpardiso->n, 2 * size), &mpardiso->schur_work_size));
  PetscCall(PetscMalloc1(mpardiso->schur_work_size, &mpardiso->schur_work));
  PetscCall(MatDestroy(&F->schur));
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, size, size, NULL, &F->schur));
  PetscCall(MatDenseGetArrayRead(F->schur, &arr));
  mpardiso->schur      = (PetscScalar *)arr;
  mpardiso->schur_size = size;
  PetscCall(MatDenseRestoreArrayRead(F->schur, &arr));
  if (mpardiso->mtype == 2) PetscCall(MatSetOption(F->schur, MAT_SPD, PETSC_TRUE));

  PetscCall(PetscFree(mpardiso->schur_idxs));
  PetscCall(PetscMalloc1(size, &mpardiso->schur_idxs));
  PetscCall(PetscArrayzero(mpardiso->perm, mpardiso->n));
  PetscCall(ISGetIndices(is, &idxs));
  PetscCall(PetscArraycpy(mpardiso->schur_idxs, idxs, size));
  for (i = 0; i < size; i++) mpardiso->perm[idxs[i]] = 1;
  PetscCall(ISRestoreIndices(is, &idxs));
  if (size) { /* turn on Schur switch if the set of indices is not empty */
    mpardiso->iparm[36 - 1] = 2;
  }
  PetscFunctionReturn(0);
}

PetscErrorCode MatDestroy_MKL_PARDISO(Mat A) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;

  PetscFunctionBegin;
  if (mat_mkl_pardiso->CleanUp) {
    mat_mkl_pardiso->phase = JOB_RELEASE_OF_ALL_MEMORY;

    MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, NULL, NULL, NULL, NULL, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, NULL, NULL,
                &mat_mkl_pardiso->err);
  }
  PetscCall(PetscFree(mat_mkl_pardiso->perm));
  PetscCall(PetscFree(mat_mkl_pardiso->schur_work));
  PetscCall(PetscFree(mat_mkl_pardiso->schur_idxs));
  if (mat_mkl_pardiso->freeaij) {
    PetscCall(PetscFree2(mat_mkl_pardiso->ia, mat_mkl_pardiso->ja));
    if (mat_mkl_pardiso->iparm[34] == 1) PetscCall(PetscFree(mat_mkl_pardiso->a));
  }
  PetscCall(PetscFree(A->data));

  /* clear composed functions */
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorSetSchurIS_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMkl_PardisoSetCntl_C", NULL));
  PetscFunctionReturn(0);
}

static PetscErrorCode MatMKLPardisoScatterSchur_Private(Mat_MKL_PARDISO *mpardiso, PetscScalar *whole, PetscScalar *schur, PetscBool reduce) {
  PetscFunctionBegin;
  if (reduce) { /* data given for the whole matrix */
    PetscInt i, m = 0, p = 0;
    for (i = 0; i < mpardiso->nrhs; i++) {
      PetscInt j;
      for (j = 0; j < mpardiso->schur_size; j++) schur[p + j] = whole[m + mpardiso->schur_idxs[j]];
      m += mpardiso->n;
      p += mpardiso->schur_size;
    }
  } else { /* from Schur to whole */
    PetscInt i, m = 0, p = 0;
    for (i = 0; i < mpardiso->nrhs; i++) {
      PetscInt j;
      for (j = 0; j < mpardiso->schur_size; j++) whole[m + mpardiso->schur_idxs[j]] = schur[p + j];
      m += mpardiso->n;
      p += mpardiso->schur_size;
    }
  }
  PetscFunctionReturn(0);
}

PetscErrorCode MatSolve_MKL_PARDISO(Mat A, Vec b, Vec x) {
  Mat_MKL_PARDISO   *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
  PetscScalar       *xarray;
  const PetscScalar *barray;

  PetscFunctionBegin;
  mat_mkl_pardiso->nrhs = 1;
  PetscCall(VecGetArrayWrite(x, &xarray));
  PetscCall(VecGetArrayRead(b, &barray));

  if (!mat_mkl_pardiso->schur) mat_mkl_pardiso->phase = JOB_SOLVE_ITERATIVE_REFINEMENT;
  else mat_mkl_pardiso->phase = JOB_SOLVE_FORWARD_SUBSTITUTION;

  if (barray == xarray) { /* if the two vectors share the same memory */
    PetscScalar *work;
    if (!mat_mkl_pardiso->schur_work) {
      PetscCall(PetscMalloc1(mat_mkl_pardiso->n, &work));
    } else {
      work = mat_mkl_pardiso->schur_work;
    }
    mat_mkl_pardiso->iparm[6 - 1] = 1;
    MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, NULL, &mat_mkl_pardiso->nrhs,
                mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)work, &mat_mkl_pardiso->err);
    if (!mat_mkl_pardiso->schur_work) PetscCall(PetscFree(work));
  } else {
    mat_mkl_pardiso->iparm[6 - 1] = 0;
    MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
                &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err);
  }
  PetscCall(VecRestoreArrayRead(b, &barray));

  PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);

  if (mat_mkl_pardiso->schur) { /* solve Schur complement and expand solution */
    if (!mat_mkl_pardiso->solve_interior) {
      PetscInt shift = mat_mkl_pardiso->schur_size;

      PetscCall(MatFactorFactorizeSchurComplement(A));
      /* if inverted, uses BLAS *MM subroutines, otherwise LAPACK *TRS */
      if (A->schur_status != MAT_FACTOR_SCHUR_INVERTED) shift = 0;

      /* solve Schur complement */
      PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work, PETSC_TRUE));
      PetscCall(MatMKLPardisoSolveSchur_Private(A, mat_mkl_pardiso->schur_work, mat_mkl_pardiso->schur_work + shift));
      PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work + shift, PETSC_FALSE));
    } else { /* if we are solving for the interior problem, any value in barray[schur] forward-substituted to xarray[schur] will be neglected */
      PetscInt i;
      for (i = 0; i < mat_mkl_pardiso->schur_size; i++) xarray[mat_mkl_pardiso->schur_idxs[i]] = 0.;
    }

    /* expansion phase */
    mat_mkl_pardiso->iparm[6 - 1] = 1;
    mat_mkl_pardiso->phase        = JOB_SOLVE_BACKWARD_SUBSTITUTION;
    MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
                &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)mat_mkl_pardiso->schur_work, /* according to the specs, the solution vector is always used */
                &mat_mkl_pardiso->err);

    PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);
    mat_mkl_pardiso->iparm[6 - 1] = 0;
  }
  PetscCall(VecRestoreArrayWrite(x, &xarray));
  mat_mkl_pardiso->CleanUp = PETSC_TRUE;
  PetscFunctionReturn(0);
}

PetscErrorCode MatSolveTranspose_MKL_PARDISO(Mat A, Vec b, Vec x) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
  PetscInt         oiparm12;

  PetscFunctionBegin;
  oiparm12                       = mat_mkl_pardiso->iparm[12 - 1];
  mat_mkl_pardiso->iparm[12 - 1] = 2;
  PetscCall(MatSolve_MKL_PARDISO(A, b, x));
  mat_mkl_pardiso->iparm[12 - 1] = oiparm12;
  PetscFunctionReturn(0);
}

PetscErrorCode MatMatSolve_MKL_PARDISO(Mat A, Mat B, Mat X) {
  Mat_MKL_PARDISO   *mat_mkl_pardiso = (Mat_MKL_PARDISO *)(A)->data;
  const PetscScalar *barray;
  PetscScalar       *xarray;
  PetscBool          flg;

  PetscFunctionBegin;
  PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQDENSE, &flg));
  PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Matrix B must be MATSEQDENSE matrix");
  if (X != B) {
    PetscCall(PetscObjectBaseTypeCompare((PetscObject)X, MATSEQDENSE, &flg));
    PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Matrix X must be MATSEQDENSE matrix");
  }

  PetscCall(MatGetSize(B, NULL, (PetscInt *)&mat_mkl_pardiso->nrhs));

  if (mat_mkl_pardiso->nrhs > 0) {
    PetscCall(MatDenseGetArrayRead(B, &barray));
    PetscCall(MatDenseGetArrayWrite(X, &xarray));

    PetscCheck(barray != xarray, PETSC_COMM_SELF, PETSC_ERR_SUP, "B and X cannot share the same memory location");
    if (!mat_mkl_pardiso->schur) mat_mkl_pardiso->phase = JOB_SOLVE_ITERATIVE_REFINEMENT;
    else mat_mkl_pardiso->phase = JOB_SOLVE_FORWARD_SUBSTITUTION;

    MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
                &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err);
    PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);

    PetscCall(MatDenseRestoreArrayRead(B, &barray));
    if (mat_mkl_pardiso->schur) { /* solve Schur complement and expand solution */
      PetscScalar *o_schur_work = NULL;

      /* solve Schur complement */
      if (!mat_mkl_pardiso->solve_interior) {
        PetscInt shift = mat_mkl_pardiso->schur_size * mat_mkl_pardiso->nrhs, scale;
        PetscInt mem   = mat_mkl_pardiso->n * mat_mkl_pardiso->nrhs;

        PetscCall(MatFactorFactorizeSchurComplement(A));
        /* allocate extra memory if it is needed */
        scale = 1;
        if (A->schur_status == MAT_FACTOR_SCHUR_INVERTED) scale = 2;
        mem *= scale;
        if (mem > mat_mkl_pardiso->schur_work_size) {
          o_schur_work = mat_mkl_pardiso->schur_work;
          PetscCall(PetscMalloc1(mem, &mat_mkl_pardiso->schur_work));
        }
        /* if inverted, uses BLAS *MM subroutines, otherwise LAPACK *TRS */
        if (A->schur_status != MAT_FACTOR_SCHUR_INVERTED) shift = 0;
        PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work, PETSC_TRUE));
        PetscCall(MatMKLPardisoSolveSchur_Private(A, mat_mkl_pardiso->schur_work, mat_mkl_pardiso->schur_work + shift));
        PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work + shift, PETSC_FALSE));
      } else { /* if we are solving for the interior problem, any value in barray[schur,n] forward-substituted to xarray[schur,n] will be neglected */
        PetscInt i, n, m = 0;
        for (n = 0; n < mat_mkl_pardiso->nrhs; n++) {
          for (i = 0; i < mat_mkl_pardiso->schur_size; i++) xarray[mat_mkl_pardiso->schur_idxs[i] + m] = 0.;
          m += mat_mkl_pardiso->n;
        }
      }

      /* expansion phase */
      mat_mkl_pardiso->iparm[6 - 1] = 1;
      mat_mkl_pardiso->phase        = JOB_SOLVE_BACKWARD_SUBSTITUTION;
      MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
                  &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)mat_mkl_pardiso->schur_work, /* according to the specs, the solution vector is always used */
                  &mat_mkl_pardiso->err);
      if (o_schur_work) { /* restore original schur_work (minimal size) */
        PetscCall(PetscFree(mat_mkl_pardiso->schur_work));
        mat_mkl_pardiso->schur_work = o_schur_work;
      }
      PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);
      mat_mkl_pardiso->iparm[6 - 1] = 0;
    }
    PetscCall(MatDenseRestoreArrayWrite(X, &xarray));
  }
  mat_mkl_pardiso->CleanUp = PETSC_TRUE;
  PetscFunctionReturn(0);
}

PetscErrorCode MatFactorNumeric_MKL_PARDISO(Mat F, Mat A, const MatFactorInfo *info) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)(F)->data;

  PetscFunctionBegin;
  mat_mkl_pardiso->matstruc = SAME_NONZERO_PATTERN;
  PetscCall((*mat_mkl_pardiso->Convert)(A, mat_mkl_pardiso->needsym, MAT_REUSE_MATRIX, &mat_mkl_pardiso->freeaij, &mat_mkl_pardiso->nz, &mat_mkl_pardiso->ia, &mat_mkl_pardiso->ja, (PetscScalar **)&mat_mkl_pardiso->a));

  mat_mkl_pardiso->phase = JOB_NUMERICAL_FACTORIZATION;
  MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
              &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, NULL, (void *)mat_mkl_pardiso->schur, &mat_mkl_pardiso->err);
  PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);

  /* report flops */
  if (mat_mkl_pardiso->iparm[18] > 0) PetscCall(PetscLogFlops(PetscPowRealInt(10., 6) * mat_mkl_pardiso->iparm[18]));

  if (F->schur) { /* schur output from pardiso is in row major format */
#if defined(PETSC_HAVE_CUDA)
    F->schur->offloadmask = PETSC_OFFLOAD_CPU;
#endif
    PetscCall(MatFactorRestoreSchurComplement(F, NULL, MAT_FACTOR_SCHUR_UNFACTORED));
    PetscCall(MatTranspose(F->schur, MAT_INPLACE_MATRIX, &F->schur));
  }
  mat_mkl_pardiso->matstruc = SAME_NONZERO_PATTERN;
  mat_mkl_pardiso->CleanUp  = PETSC_TRUE;
  PetscFunctionReturn(0);
}

PetscErrorCode MatSetFromOptions_MKL_PARDISO(Mat F, Mat A) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
  PetscInt         icntl, bs, threads = 1;
  PetscBool        flg;

  PetscFunctionBegin;
  PetscOptionsBegin(PetscObjectComm((PetscObject)F), ((PetscObject)F)->prefix, "MKL_PARDISO Options", "Mat");

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_65", "Suggested number of threads to use within PARDISO", "None", threads, &threads, &flg));
  if (flg) PetscSetMKL_PARDISOThreads((int)threads);

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_66", "Maximum number of factors with identical sparsity structure that must be kept in memory at the same time", "None", mat_mkl_pardiso->maxfct, &icntl, &flg));
  if (flg) mat_mkl_pardiso->maxfct = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_67", "Indicates the actual matrix for the solution phase", "None", mat_mkl_pardiso->mnum, &icntl, &flg));
  if (flg) mat_mkl_pardiso->mnum = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_68", "Message level information", "None", mat_mkl_pardiso->msglvl, &icntl, &flg));
  if (flg) mat_mkl_pardiso->msglvl = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_69", "Defines the matrix type", "None", mat_mkl_pardiso->mtype, &icntl, &flg));
  if (flg) {
    void *pt[IPARM_SIZE];
    mat_mkl_pardiso->mtype = icntl;
    icntl                  = mat_mkl_pardiso->iparm[34];
    bs                     = mat_mkl_pardiso->iparm[36];
    MKL_PARDISO_INIT(pt, &mat_mkl_pardiso->mtype, mat_mkl_pardiso->iparm);
#if defined(PETSC_USE_REAL_SINGLE)
    mat_mkl_pardiso->iparm[27] = 1;
#else
    mat_mkl_pardiso->iparm[27] = 0;
#endif
    mat_mkl_pardiso->iparm[34] = icntl;
    mat_mkl_pardiso->iparm[36] = bs;
  }

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_1", "Use default values (if 0)", "None", mat_mkl_pardiso->iparm[0], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[0] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_2", "Fill-in reducing ordering for the input matrix", "None", mat_mkl_pardiso->iparm[1], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[1] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_4", "Preconditioned CGS/CG", "None", mat_mkl_pardiso->iparm[3], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[3] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_5", "User permutation", "None", mat_mkl_pardiso->iparm[4], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[4] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_6", "Write solution on x", "None", mat_mkl_pardiso->iparm[5], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[5] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_8", "Iterative refinement step", "None", mat_mkl_pardiso->iparm[7], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[7] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_10", "Pivoting perturbation", "None", mat_mkl_pardiso->iparm[9], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[9] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_11", "Scaling vectors", "None", mat_mkl_pardiso->iparm[10], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[10] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_12", "Solve with transposed or conjugate transposed matrix A", "None", mat_mkl_pardiso->iparm[11], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[11] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_13", "Improved accuracy using (non-) symmetric weighted matching", "None", mat_mkl_pardiso->iparm[12], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[12] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_18", "Numbers of non-zero elements", "None", mat_mkl_pardiso->iparm[17], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[17] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_19", "Report number of floating point operations (0 to disable)", "None", mat_mkl_pardiso->iparm[18], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[18] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_21", "Pivoting for symmetric indefinite matrices", "None", mat_mkl_pardiso->iparm[20], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[20] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_24", "Parallel factorization control", "None", mat_mkl_pardiso->iparm[23], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[23] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_25", "Parallel forward/backward solve control", "None", mat_mkl_pardiso->iparm[24], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[24] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_27", "Matrix checker", "None", mat_mkl_pardiso->iparm[26], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[26] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_31", "Partial solve and computing selected components of the solution vectors", "None", mat_mkl_pardiso->iparm[30], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[30] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_34", "Optimal number of threads for conditional numerical reproducibility (CNR) mode", "None", mat_mkl_pardiso->iparm[33], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[33] = icntl;

  PetscCall(PetscOptionsInt("-mat_mkl_pardiso_60", "Intel MKL_PARDISO mode", "None", mat_mkl_pardiso->iparm[59], &icntl, &flg));
  if (flg) mat_mkl_pardiso->iparm[59] = icntl;
  PetscOptionsEnd();
  PetscFunctionReturn(0);
}

PetscErrorCode MatFactorMKL_PARDISOInitialize_Private(Mat A, MatFactorType ftype, Mat_MKL_PARDISO *mat_mkl_pardiso) {
  PetscInt  i, bs;
  PetscBool match;

  PetscFunctionBegin;
  for (i = 0; i < IPARM_SIZE; i++) mat_mkl_pardiso->iparm[i] = 0;
  for (i = 0; i < IPARM_SIZE; i++) mat_mkl_pardiso->pt[i] = 0;
#if defined(PETSC_USE_REAL_SINGLE)
  mat_mkl_pardiso->iparm[27] = 1;
#else
  mat_mkl_pardiso->iparm[27] = 0;
#endif
  /* Default options for both sym and unsym */
  mat_mkl_pardiso->iparm[0]  = 1;  /* Solver default parameters overriden with provided by iparm */
  mat_mkl_pardiso->iparm[1]  = 2;  /* Metis reordering */
  mat_mkl_pardiso->iparm[5]  = 0;  /* Write solution into x */
  mat_mkl_pardiso->iparm[7]  = 0;  /* Max number of iterative refinement steps */
  mat_mkl_pardiso->iparm[17] = -1; /* Output: Number of nonzeros in the factor LU */
  mat_mkl_pardiso->iparm[18] = -1; /* Output: Mflops for LU factorization */
#if 0
  mat_mkl_pardiso->iparm[23] =  1; /* Parallel factorization control*/
#endif
  PetscCall(PetscObjectTypeCompareAny((PetscObject)A, &match, MATSEQBAIJ, MATSEQSBAIJ, ""));
  PetscCall(MatGetBlockSize(A, &bs));
  if (!match || bs == 1) {
    mat_mkl_pardiso->iparm[34] = 1; /* Cluster Sparse Solver use C-style indexing for ia and ja arrays */
    mat_mkl_pardiso->n         = A->rmap->N;
  } else {
    mat_mkl_pardiso->iparm[34] = 0; /* Cluster Sparse Solver use Fortran-style indexing for ia and ja arrays */
    mat_mkl_pardiso->iparm[36] = bs;
    mat_mkl_pardiso->n         = A->rmap->N / bs;
  }
  mat_mkl_pardiso->iparm[39] = 0; /* Input: matrix/rhs/solution stored on rank-0 */

  mat_mkl_pardiso->CleanUp = PETSC_FALSE;
  mat_mkl_pardiso->maxfct  = 1; /* Maximum number of numerical factorizations. */
  mat_mkl_pardiso->mnum    = 1; /* Which factorization to use. */
  mat_mkl_pardiso->msglvl  = 0; /* 0: do not print 1: Print statistical information in file */
  mat_mkl_pardiso->phase   = -1;
  mat_mkl_pardiso->err     = 0;

  mat_mkl_pardiso->nrhs  = 1;
  mat_mkl_pardiso->err   = 0;
  mat_mkl_pardiso->phase = -1;

  if (ftype == MAT_FACTOR_LU) {
    mat_mkl_pardiso->iparm[9]  = 13; /* Perturb the pivot elements with 1E-13 */
    mat_mkl_pardiso->iparm[10] = 1;  /* Use nonsymmetric permutation and scaling MPS */
    mat_mkl_pardiso->iparm[12] = 1;  /* Switch on Maximum Weighted Matching algorithm (default for non-symmetric) */
  } else {
    mat_mkl_pardiso->iparm[9]  = 8; /* Perturb the pivot elements with 1E-8 */
    mat_mkl_pardiso->iparm[10] = 0; /* Use nonsymmetric permutation and scaling MPS */
    mat_mkl_pardiso->iparm[12] = 1; /* Switch on Maximum Weighted Matching algorithm (default for non-symmetric) */
#if defined(PETSC_USE_DEBUG)
    mat_mkl_pardiso->iparm[26] = 1; /* Matrix checker */
#endif
  }
  PetscCall(PetscCalloc1(A->rmap->N * sizeof(INT_TYPE), &mat_mkl_pardiso->perm));
  mat_mkl_pardiso->schur_size = 0;
  PetscFunctionReturn(0);
}

PetscErrorCode MatFactorSymbolic_AIJMKL_PARDISO_Private(Mat F, Mat A, const MatFactorInfo *info) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;

  PetscFunctionBegin;
  mat_mkl_pardiso->matstruc = DIFFERENT_NONZERO_PATTERN;
  PetscCall(MatSetFromOptions_MKL_PARDISO(F, A));
  /* throw away any previously computed structure */
  if (mat_mkl_pardiso->freeaij) {
    PetscCall(PetscFree2(mat_mkl_pardiso->ia, mat_mkl_pardiso->ja));
    if (mat_mkl_pardiso->iparm[34] == 1) PetscCall(PetscFree(mat_mkl_pardiso->a));
  }
  PetscCall((*mat_mkl_pardiso->Convert)(A, mat_mkl_pardiso->needsym, MAT_INITIAL_MATRIX, &mat_mkl_pardiso->freeaij, &mat_mkl_pardiso->nz, &mat_mkl_pardiso->ia, &mat_mkl_pardiso->ja, (PetscScalar **)&mat_mkl_pardiso->a));
  if (mat_mkl_pardiso->iparm[34] == 1) mat_mkl_pardiso->n = A->rmap->N;
  else mat_mkl_pardiso->n = A->rmap->N / A->rmap->bs;

  mat_mkl_pardiso->phase = JOB_ANALYSIS;

  /* reset flops counting if requested */
  if (mat_mkl_pardiso->iparm[18]) mat_mkl_pardiso->iparm[18] = -1;

  MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
              &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, NULL, NULL, &mat_mkl_pardiso->err);
  PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL_PARDISO: err=%d. Please check manual", mat_mkl_pardiso->err);

  mat_mkl_pardiso->CleanUp = PETSC_TRUE;

  if (F->factortype == MAT_FACTOR_LU) F->ops->lufactornumeric = MatFactorNumeric_MKL_PARDISO;
  else F->ops->choleskyfactornumeric = MatFactorNumeric_MKL_PARDISO;

  F->ops->solve          = MatSolve_MKL_PARDISO;
  F->ops->solvetranspose = MatSolveTranspose_MKL_PARDISO;
  F->ops->matsolve       = MatMatSolve_MKL_PARDISO;
  PetscFunctionReturn(0);
}

PetscErrorCode MatLUFactorSymbolic_AIJMKL_PARDISO(Mat F, Mat A, IS r, IS c, const MatFactorInfo *info) {
  PetscFunctionBegin;
  PetscCall(MatFactorSymbolic_AIJMKL_PARDISO_Private(F, A, info));
  PetscFunctionReturn(0);
}

#if !defined(PETSC_USE_COMPLEX)
PetscErrorCode MatGetInertia_MKL_PARDISO(Mat F, PetscInt *nneg, PetscInt *nzero, PetscInt *npos) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;

  PetscFunctionBegin;
  if (nneg) *nneg = mat_mkl_pardiso->iparm[22];
  if (npos) *npos = mat_mkl_pardiso->iparm[21];
  if (nzero) *nzero = F->rmap->N - (mat_mkl_pardiso->iparm[22] + mat_mkl_pardiso->iparm[21]);
  PetscFunctionReturn(0);
}
#endif

PetscErrorCode MatCholeskyFactorSymbolic_AIJMKL_PARDISO(Mat F, Mat A, IS r, const MatFactorInfo *info) {
  PetscFunctionBegin;
  PetscCall(MatFactorSymbolic_AIJMKL_PARDISO_Private(F, A, info));
#if defined(PETSC_USE_COMPLEX)
  F->ops->getinertia = NULL;
#else
  F->ops->getinertia = MatGetInertia_MKL_PARDISO;
#endif
  PetscFunctionReturn(0);
}

PetscErrorCode MatView_MKL_PARDISO(Mat A, PetscViewer viewer) {
  PetscBool         iascii;
  PetscViewerFormat format;
  Mat_MKL_PARDISO  *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
  PetscInt          i;

  PetscFunctionBegin;
  if (A->ops->solve != MatSolve_MKL_PARDISO) PetscFunctionReturn(0);

  PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
  if (iascii) {
    PetscCall(PetscViewerGetFormat(viewer, &format));
    if (format == PETSC_VIEWER_ASCII_INFO) {
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO run parameters:\n"));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO phase:             %d \n", mat_mkl_pardiso->phase));
      for (i = 1; i <= 64; i++) PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO iparm[%d]:     %d \n", i, mat_mkl_pardiso->iparm[i - 1]));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO maxfct:     %d \n", mat_mkl_pardiso->maxfct));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO mnum:     %d \n", mat_mkl_pardiso->mnum));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO mtype:     %d \n", mat_mkl_pardiso->mtype));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO n:     %d \n", mat_mkl_pardiso->n));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO nrhs:     %d \n", mat_mkl_pardiso->nrhs));
      PetscCall(PetscViewerASCIIPrintf(viewer, "MKL_PARDISO msglvl:     %d \n", mat_mkl_pardiso->msglvl));
    }
  }
  PetscFunctionReturn(0);
}

PetscErrorCode MatGetInfo_MKL_PARDISO(Mat A, MatInfoType flag, MatInfo *info) {
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;

  PetscFunctionBegin;
  info->block_size        = 1.0;
  info->nz_used           = mat_mkl_pardiso->iparm[17];
  info->nz_allocated      = mat_mkl_pardiso->iparm[17];
  info->nz_unneeded       = 0.0;
  info->assemblies        = 0.0;
  info->mallocs           = 0.0;
  info->memory            = 0.0;
  info->fill_ratio_given  = 0;
  info->fill_ratio_needed = 0;
  info->factor_mallocs    = 0;
  PetscFunctionReturn(0);
}

PetscErrorCode MatMkl_PardisoSetCntl_MKL_PARDISO(Mat F, PetscInt icntl, PetscInt ival) {
  PetscInt         backup, bs;
  Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;

  PetscFunctionBegin;
  if (icntl <= 64) {
    mat_mkl_pardiso->iparm[icntl - 1] = ival;
  } else {
    if (icntl == 65) PetscSetMKL_PARDISOThreads(ival);
    else if (icntl == 66) mat_mkl_pardiso->maxfct = ival;
    else if (icntl == 67) mat_mkl_pardiso->mnum = ival;
    else if (icntl == 68) mat_mkl_pardiso->msglvl = ival;
    else if (icntl == 69) {
      void *pt[IPARM_SIZE];
      backup                 = mat_mkl_pardiso->iparm[34];
      bs                     = mat_mkl_pardiso->iparm[36];
      mat_mkl_pardiso->mtype = ival;
      MKL_PARDISO_INIT(pt, &mat_mkl_pardiso->mtype, mat_mkl_pardiso->iparm);
#if defined(PETSC_USE_REAL_SINGLE)
      mat_mkl_pardiso->iparm[27] = 1;
#else
      mat_mkl_pardiso->iparm[27] = 0;
#endif
      mat_mkl_pardiso->iparm[34] = backup;
      mat_mkl_pardiso->iparm[36] = bs;
    } else if (icntl == 70) mat_mkl_pardiso->solve_interior = (PetscBool) !!ival;
  }
  PetscFunctionReturn(0);
}

/*@
  MatMkl_PardisoSetCntl - Set Mkl_Pardiso parameters

   Logically Collective on Mat

   Input Parameters:
+  F - the factored matrix obtained by calling MatGetFactor()
.  icntl - index of Mkl_Pardiso parameter
-  ival - value of Mkl_Pardiso parameter

  Options Database:
.   -mat_mkl_pardiso_<icntl> <ival> - change the option numbered icntl to the value ival

   Level: beginner

   References:
.  * - Mkl_Pardiso Users' Guide

.seealso: `MatGetFactor()`
@*/
PetscErrorCode MatMkl_PardisoSetCntl(Mat F, PetscInt icntl, PetscInt ival) {
  PetscFunctionBegin;
  PetscTryMethod(F, "MatMkl_PardisoSetCntl_C", (Mat, PetscInt, PetscInt), (F, icntl, ival));
  PetscFunctionReturn(0);
}

/*MC
  MATSOLVERMKL_PARDISO -  A matrix type providing direct solvers (LU) for
  sequential matrices via the external package MKL_PARDISO.

  Works with MATSEQAIJ matrices

  Use -pc_type lu -pc_factor_mat_solver_type mkl_pardiso to use this direct solver

  Options Database Keys:
+ -mat_mkl_pardiso_65 - Suggested number of threads to use within MKL_PARDISO
. -mat_mkl_pardiso_66 - Maximum number of factors with identical sparsity structure that must be kept in memory at the same time
. -mat_mkl_pardiso_67 - Indicates the actual matrix for the solution phase
. -mat_mkl_pardiso_68 - Message level information, use 1 to get detailed information on the solver options
. -mat_mkl_pardiso_69 - Defines the matrix type. IMPORTANT: When you set this flag, iparm parameters are going to be set to the default ones for the matrix type
. -mat_mkl_pardiso_1  - Use default values
. -mat_mkl_pardiso_2  - Fill-in reducing ordering for the input matrix
. -mat_mkl_pardiso_4  - Preconditioned CGS/CG
. -mat_mkl_pardiso_5  - User permutation
. -mat_mkl_pardiso_6  - Write solution on x
. -mat_mkl_pardiso_8  - Iterative refinement step
. -mat_mkl_pardiso_10 - Pivoting perturbation
. -mat_mkl_pardiso_11 - Scaling vectors
. -mat_mkl_pardiso_12 - Solve with transposed or conjugate transposed matrix A
. -mat_mkl_pardiso_13 - Improved accuracy using (non-) symmetric weighted matching
. -mat_mkl_pardiso_18 - Numbers of non-zero elements
. -mat_mkl_pardiso_19 - Report number of floating point operations
. -mat_mkl_pardiso_21 - Pivoting for symmetric indefinite matrices
. -mat_mkl_pardiso_24 - Parallel factorization control
. -mat_mkl_pardiso_25 - Parallel forward/backward solve control
. -mat_mkl_pardiso_27 - Matrix checker
. -mat_mkl_pardiso_31 - Partial solve and computing selected components of the solution vectors
. -mat_mkl_pardiso_34 - Optimal number of threads for conditional numerical reproducibility (CNR) mode
- -mat_mkl_pardiso_60 - Intel MKL_PARDISO mode

  Level: beginner

  Notes:
    Use -mat_mkl_pardiso_68 1 to display the number of threads the solver is using. MKL does not provide a way to directly access this
    information.

    For more information on the options check the MKL_Pardiso manual

.seealso: `PCFactorSetMatSolverType()`, `MatSolverType`

M*/
static PetscErrorCode MatFactorGetSolverType_mkl_pardiso(Mat A, MatSolverType *type) {
  PetscFunctionBegin;
  *type = MATSOLVERMKL_PARDISO;
  PetscFunctionReturn(0);
}

PETSC_EXTERN PetscErrorCode MatGetFactor_aij_mkl_pardiso(Mat A, MatFactorType ftype, Mat *F) {
  Mat              B;
  Mat_MKL_PARDISO *mat_mkl_pardiso;
  PetscBool        isSeqAIJ, isSeqBAIJ, isSeqSBAIJ;

  PetscFunctionBegin;
  PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isSeqAIJ));
  PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQBAIJ, &isSeqBAIJ));
  PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQSBAIJ, &isSeqSBAIJ));
  PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
  PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
  PetscCall(PetscStrallocpy("mkl_pardiso", &((PetscObject)B)->type_name));
  PetscCall(MatSetUp(B));

  PetscCall(PetscNewLog(B, &mat_mkl_pardiso));
  B->data = mat_mkl_pardiso;

  PetscCall(MatFactorMKL_PARDISOInitialize_Private(A, ftype, mat_mkl_pardiso));
  if (ftype == MAT_FACTOR_LU) {
    B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMKL_PARDISO;
    B->factortype            = MAT_FACTOR_LU;
    mat_mkl_pardiso->needsym = PETSC_FALSE;
    if (isSeqAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqaij;
    else if (isSeqBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqbaij;
    else {
      PetscCheck(!isSeqSBAIJ, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for PARDISO LU factor with SEQSBAIJ format! Use MAT_FACTOR_CHOLESKY instead");
      SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for PARDISO LU with %s format", ((PetscObject)A)->type_name);
    }
#if defined(PETSC_USE_COMPLEX)
    mat_mkl_pardiso->mtype = 13;
#else
    mat_mkl_pardiso->mtype = 11;
#endif
  } else {
    B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_AIJMKL_PARDISO;
    B->factortype                  = MAT_FACTOR_CHOLESKY;
    if (isSeqAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqaij;
    else if (isSeqBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqbaij;
    else if (isSeqSBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqsbaij;
    else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for PARDISO CHOLESKY with %s format", ((PetscObject)A)->type_name);

    mat_mkl_pardiso->needsym = PETSC_TRUE;
#if !defined(PETSC_USE_COMPLEX)
    if (A->spd == PETSC_BOOL3_TRUE) mat_mkl_pardiso->mtype = 2;
    else mat_mkl_pardiso->mtype = -2;
#else
    mat_mkl_pardiso->mtype = 6;
    PetscCheck(A->hermitian != PETSC_BOOL3_TRUE, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for PARDISO CHOLESKY with Hermitian matrices! Use MAT_FACTOR_LU instead");
#endif
  }
  B->ops->destroy = MatDestroy_MKL_PARDISO;
  B->ops->view    = MatView_MKL_PARDISO;
  B->ops->getinfo = MatGetInfo_MKL_PARDISO;
  B->factortype   = ftype;
  B->assembled    = PETSC_TRUE;

  PetscCall(PetscFree(B->solvertype));
  PetscCall(PetscStrallocpy(MATSOLVERMKL_PARDISO, &B->solvertype));

  PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mkl_pardiso));
  PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MKL_PARDISO));
  PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMkl_PardisoSetCntl_C", MatMkl_PardisoSetCntl_MKL_PARDISO));

  *F = B;
  PetscFunctionReturn(0);
}

PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_MKL_Pardiso(void) {
  PetscFunctionBegin;
  PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mkl_pardiso));
  PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mkl_pardiso));
  PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQBAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mkl_pardiso));
  PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQSBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mkl_pardiso));
  PetscFunctionReturn(0);
}