seq/aijmkl/aijmkl.c

/*
  Defines basic operations for the MATSEQAIJMKL matrix class.
  This class is derived from the MATSEQAIJ class and retains the
  compressed row storage (aka Yale sparse matrix format) but uses
  sparse BLAS operations from the Intel Math Kernel Library (MKL)
  wherever possible.
*/

#include <../src/mat/impls/aij/seq/aij.h>
#include <../src/mat/impls/aij/seq/aijmkl/aijmkl.h>
#include <mkl_spblas.h>

typedef struct {
  PetscBool        no_SpMV2;         /* If PETSC_TRUE, then don't use the MKL SpMV2 inspector-executor routines. */
  PetscBool        eager_inspection; /* If PETSC_TRUE, then call mkl_sparse_optimize() in MatDuplicate()/MatAssemblyEnd(). */
  PetscBool        sparse_optimized; /* If PETSC_TRUE, then mkl_sparse_optimize() has been called. */
  PetscObjectState state;
#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  sparse_matrix_t     csrA; /* "Handle" used by SpMV2 inspector-executor routines. */
  struct matrix_descr descr;
#endif
} Mat_SeqAIJMKL;

extern PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat, MatAssemblyType);

PETSC_INTERN PetscErrorCode MatConvert_SeqAIJMKL_SeqAIJ(Mat A, MatType type, MatReuse reuse, Mat *newmat) {
  /* This routine is only called to convert a MATAIJMKL to its base PETSc type, */
  /* so we will ignore 'MatType type'. */
  Mat B = *newmat;
#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
#endif

  PetscFunctionBegin;
  if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &B));

  /* Reset the original function pointers. */
  B->ops->duplicate               = MatDuplicate_SeqAIJ;
  B->ops->assemblyend             = MatAssemblyEnd_SeqAIJ;
  B->ops->destroy                 = MatDestroy_SeqAIJ;
  B->ops->mult                    = MatMult_SeqAIJ;
  B->ops->multtranspose           = MatMultTranspose_SeqAIJ;
  B->ops->multadd                 = MatMultAdd_SeqAIJ;
  B->ops->multtransposeadd        = MatMultTransposeAdd_SeqAIJ;
  B->ops->productsetfromoptions   = MatProductSetFromOptions_SeqAIJ;
  B->ops->matmultsymbolic         = MatMatMultSymbolic_SeqAIJ_SeqAIJ;
  B->ops->matmultnumeric          = MatMatMultNumeric_SeqAIJ_SeqAIJ;
  B->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ;
  B->ops->transposematmultnumeric = MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ;
  B->ops->ptapnumeric             = MatPtAPNumeric_SeqAIJ_SeqAIJ;

  PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaijmkl_seqaij_C", NULL));

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  /* Free everything in the Mat_SeqAIJMKL data structure. Currently, this
   * simply involves destroying the MKL sparse matrix handle and then freeing
   * the spptr pointer. */
  if (reuse == MAT_INITIAL_MATRIX) aijmkl = (Mat_SeqAIJMKL *)B->spptr;

  if (aijmkl->sparse_optimized) PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */
  PetscCall(PetscFree(B->spptr));

  /* Change the type of B to MATSEQAIJ. */
  PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));

  *newmat = B;
  PetscFunctionReturn(0);
}

PetscErrorCode MatDestroy_SeqAIJMKL(Mat A) {
  Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;

  PetscFunctionBegin;

  /* If MatHeaderMerge() was used, then this SeqAIJMKL matrix will not have an spptr pointer. */
  if (aijmkl) {
    /* Clean up everything in the Mat_SeqAIJMKL data structure, then free A->spptr. */
#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
    if (aijmkl->sparse_optimized) PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */
    PetscCall(PetscFree(A->spptr));
  }

  /* Change the type of A back to SEQAIJ and use MatDestroy_SeqAIJ()
   * to destroy everything that remains. */
  PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATSEQAIJ));
  /* Note that I don't call MatSetType().  I believe this is because that
   * is only to be called when *building* a matrix.  I could be wrong, but
   * that is how things work for the SuperLU matrix class. */
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijmkl_seqaij_C", NULL));
  PetscCall(MatDestroy_SeqAIJ(A));
  PetscFunctionReturn(0);
}

/* MatSeqAIJMKL_create_mkl_handle(), if called with an AIJMKL matrix that has not had mkl_sparse_optimize() called for it,
 * creates an MKL sparse matrix handle from the AIJ arrays and calls mkl_sparse_optimize().
 * If called with an AIJMKL matrix for which aijmkl->sparse_optimized == PETSC_TRUE, then it destroys the old matrix
 * handle, creates a new one, and then calls mkl_sparse_optimize().
 * Although in normal MKL usage it is possible to have a valid matrix handle on which mkl_sparse_optimize() has not been
 * called, for AIJMKL the handle creation and optimization step always occur together, so we don't handle the case of
 * an unoptimized matrix handle here. */
PETSC_INTERN PetscErrorCode MatSeqAIJMKL_create_mkl_handle(Mat A) {
#if !defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  /* If the MKL library does not have mkl_sparse_optimize(), then this routine
   * does nothing. We make it callable anyway in this case because it cuts
   * down on littering the code with #ifdefs. */
  PetscFunctionBegin;
  PetscFunctionReturn(0);
#else
  Mat_SeqAIJ    *a      = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  PetscInt       m, n;
  MatScalar     *aa;
  PetscInt      *aj, *ai;

  PetscFunctionBegin;
#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
  /* For MKL versions that still support the old, non-inspector-executor interfaces versions, we simply exit here if the no_SpMV2
   * option has been specified. For versions that have deprecated the old interfaces (version 18, update 2 and later), we must
   * use the new inspector-executor interfaces, but we can still use the old, non-inspector-executor code by not calling
   * mkl_sparse_optimize() later. */
  if (aijmkl->no_SpMV2) PetscFunctionReturn(0);
#endif

  if (aijmkl->sparse_optimized) {
    /* Matrix has been previously assembled and optimized. Must destroy old
     * matrix handle before running the optimization step again. */
    PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
  }
  aijmkl->sparse_optimized = PETSC_FALSE;

  /* Now perform the SpMV2 setup and matrix optimization. */
  aijmkl->descr.type = SPARSE_MATRIX_TYPE_GENERAL;
  aijmkl->descr.mode = SPARSE_FILL_MODE_LOWER;
  aijmkl->descr.diag = SPARSE_DIAG_NON_UNIT;
  m                  = A->rmap->n;
  n                  = A->cmap->n;
  aj                 = a->j; /* aj[k] gives column index for element aa[k]. */
  aa                 = a->a; /* Nonzero elements stored row-by-row. */
  ai                 = a->i; /* ai[k] is the position in aa and aj where row k starts. */
  if (a->nz && aa && !A->structure_only) {
    /* Create a new, optimized sparse matrix handle only if the matrix has nonzero entries.
     * The MKL sparse-inspector executor routines don't like being passed an empty matrix. */
    PetscCallExternal(mkl_sparse_x_create_csr, &aijmkl->csrA, SPARSE_INDEX_BASE_ZERO, m, n, ai, ai + 1, aj, aa);
    PetscCallExternal(mkl_sparse_set_mv_hint, aijmkl->csrA, SPARSE_OPERATION_NON_TRANSPOSE, aijmkl->descr, 1000);
    PetscCallExternal(mkl_sparse_set_memory_hint, aijmkl->csrA, SPARSE_MEMORY_AGGRESSIVE);
    if (!aijmkl->no_SpMV2) PetscCallExternal(mkl_sparse_optimize, aijmkl->csrA);
    aijmkl->sparse_optimized = PETSC_TRUE;
    PetscCall(PetscObjectStateGet((PetscObject)A, &(aijmkl->state)));
  } else {
    aijmkl->csrA = PETSC_NULL;
  }

  PetscFunctionReturn(0);
#endif
}

#if defined(PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE)
/* Take an already created but empty matrix and set up the nonzero structure from an MKL sparse matrix handle. */
static PetscErrorCode MatSeqAIJMKL_setup_structure_from_mkl_handle(MPI_Comm comm, sparse_matrix_t csrA, PetscInt nrows, PetscInt ncols, Mat A) {
  sparse_index_base_t indexing;
  PetscInt            m, n;
  PetscInt           *aj, *ai, *dummy;
  MatScalar          *aa;
  Mat_SeqAIJMKL      *aijmkl;

  PetscFunctionBegin;
  if (csrA) {
    /* Note: Must pass in &dummy below since MKL can't accept NULL for this output array we don't actually want. */
    PetscCallExternal(mkl_sparse_x_export_csr, csrA, &indexing, &m, &n, &ai, &dummy, &aj, &aa);
    PetscCheck((m == nrows) && (n == ncols), PETSC_COMM_SELF, PETSC_ERR_LIB, "Number of rows/columns does not match those from mkl_sparse_x_export_csr()");
  } else {
    aj = ai = PETSC_NULL;
    aa      = PETSC_NULL;
  }

  PetscCall(MatSetType(A, MATSEQAIJ));
  PetscCall(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, nrows, ncols));
  /* We use MatSeqAIJSetPreallocationCSR() instead of MatCreateSeqAIJWithArrays() because we must copy the arrays exported
   * from MKL; MKL developers tell us that modifying the arrays may cause unexpected results when using the MKL handle, and
   * they will be destroyed when the MKL handle is destroyed.
   * (In the interest of reducing memory consumption in future, can we figure out good ways to deal with this?) */
  if (csrA) {
    PetscCall(MatSeqAIJSetPreallocationCSR(A, ai, aj, NULL));
  } else {
    /* Since MatSeqAIJSetPreallocationCSR does initial set up and assembly begin/end, we must do that ourselves here. */
    PetscCall(MatSetUp(A));
    PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
    PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
  }

  /* We now have an assembled sequential AIJ matrix created from copies of the exported arrays from the MKL matrix handle.
   * Now turn it into a MATSEQAIJMKL. */
  PetscCall(MatConvert_SeqAIJ_SeqAIJMKL(A, MATSEQAIJMKL, MAT_INPLACE_MATRIX, &A));

  aijmkl       = (Mat_SeqAIJMKL *)A->spptr;
  aijmkl->csrA = csrA;

  /* The below code duplicates much of what is in MatSeqAIJKL_create_mkl_handle(). I dislike this code duplication, but
   * MatSeqAIJMKL_create_mkl_handle() cannot be used because we don't need to create a handle -- we've already got one,
   * and just need to be able to run the MKL optimization step. */
  aijmkl->descr.type = SPARSE_MATRIX_TYPE_GENERAL;
  aijmkl->descr.mode = SPARSE_FILL_MODE_LOWER;
  aijmkl->descr.diag = SPARSE_DIAG_NON_UNIT;
  if (csrA) {
    PetscCallExternal(mkl_sparse_set_mv_hint, aijmkl->csrA, SPARSE_OPERATION_NON_TRANSPOSE, aijmkl->descr, 1000);
    PetscCallExternal(mkl_sparse_set_memory_hint, aijmkl->csrA, SPARSE_MEMORY_AGGRESSIVE);
  }
  PetscCall(PetscObjectStateGet((PetscObject)A, &(aijmkl->state)));
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE */

/* MatSeqAIJMKL_update_from_mkl_handle() updates the matrix values array from the contents of the associated MKL sparse matrix handle.
 * This is needed after mkl_sparse_sp2m() with SPARSE_STAGE_FINALIZE_MULT has been used to compute new values of the matrix in
 * MatMatMultNumeric(). */
#if defined(PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE)
static PetscErrorCode MatSeqAIJMKL_update_from_mkl_handle(Mat A) {
  PetscInt            i;
  PetscInt            nrows, ncols;
  PetscInt            nz;
  PetscInt           *ai, *aj, *dummy;
  PetscScalar        *aa;
  Mat_SeqAIJMKL      *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  sparse_index_base_t indexing;

  PetscFunctionBegin;
  /* Exit immediately in case of the MKL matrix handle being NULL; this will be the case for empty matrices (zero rows or columns). */
  if (!aijmkl->csrA) PetscFunctionReturn(0);

  /* Note: Must pass in &dummy below since MKL can't accept NULL for this output array we don't actually want. */
  PetscCallExternal(mkl_sparse_x_export_csr, aijmkl->csrA, &indexing, &nrows, &ncols, &ai, &dummy, &aj, &aa);

  /* We can't just do a copy from the arrays exported by MKL to those used for the PETSc AIJ storage, because the MKL and PETSc
   * representations differ in small ways (e.g., more explicit nonzeros per row due to preallocation). */
  for (i = 0; i < nrows; i++) {
    nz = ai[i + 1] - ai[i];
    PetscCall(MatSetValues_SeqAIJ(A, 1, &i, nz, aj + ai[i], aa + ai[i], INSERT_VALUES));
  }

  PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
  PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));

  PetscCall(PetscObjectStateGet((PetscObject)A, &(aijmkl->state)));
  /* At this point our matrix has a valid MKL handle, the contents of which match the PETSc AIJ representation.
   * The MKL handle has *not* had mkl_sparse_optimize() called on it, though -- the MKL developers have confirmed
   * that the matrix inspection/optimization step is not performed when matrix-matrix multiplication is finalized. */
  aijmkl->sparse_optimized = PETSC_FALSE;
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE */

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
PETSC_INTERN PetscErrorCode MatSeqAIJMKL_view_mkl_handle(Mat A, PetscViewer viewer) {
  PetscInt            i, j, k;
  PetscInt            nrows, ncols;
  PetscInt            nz;
  PetscInt           *ai, *aj, *dummy;
  PetscScalar        *aa;
  Mat_SeqAIJMKL      *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  sparse_index_base_t indexing;

  PetscFunctionBegin;
  PetscCall(PetscViewerASCIIPrintf(viewer, "Contents of MKL sparse matrix handle for MATSEQAIJMKL object:\n"));

  /* Exit immediately in case of the MKL matrix handle being NULL; this will be the case for empty matrices (zero rows or columns). */
  if (!aijmkl->csrA) {
    PetscCall(PetscViewerASCIIPrintf(viewer, "MKL matrix handle is NULL\n"));
    PetscFunctionReturn(0);
  }

  /* Note: Must pass in &dummy below since MKL can't accept NULL for this output array we don't actually want. */
  PetscCallExternal(mkl_sparse_x_export_csr, aijmkl->csrA, &indexing, &nrows, &ncols, &ai, &dummy, &aj, &aa);

  k = 0;
  for (i = 0; i < nrows; i++) {
    PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ": ", i));
    nz = ai[i + 1] - ai[i];
    for (j = 0; j < nz; j++) {
      if (aa) {
        PetscCall(PetscViewerASCIIPrintf(viewer, "(%" PetscInt_FMT ", %g)  ", aj[k], PetscRealPart(aa[k])));
      } else {
        PetscCall(PetscViewerASCIIPrintf(viewer, "(%" PetscInt_FMT ", NULL)", aj[k]));
      }
      k++;
    }
    PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
  }
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

PetscErrorCode MatDuplicate_SeqAIJMKL(Mat A, MatDuplicateOption op, Mat *M) {
  Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  Mat_SeqAIJMKL *aijmkl_dest;

  PetscFunctionBegin;
  PetscCall(MatDuplicate_SeqAIJ(A, op, M));
  aijmkl_dest = (Mat_SeqAIJMKL *)(*M)->spptr;
  PetscCall(PetscArraycpy(aijmkl_dest, aijmkl, 1));
  aijmkl_dest->sparse_optimized = PETSC_FALSE;
  if (aijmkl->eager_inspection) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
  PetscFunctionReturn(0);
}

PetscErrorCode MatAssemblyEnd_SeqAIJMKL(Mat A, MatAssemblyType mode) {
  Mat_SeqAIJ    *a = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL *aijmkl;

  PetscFunctionBegin;
  if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0);

  /* Since a MATSEQAIJMKL matrix is really just a MATSEQAIJ with some
   * extra information and some different methods, call the AssemblyEnd
   * routine for a MATSEQAIJ.
   * I'm not sure if this is the best way to do this, but it avoids
   * a lot of code duplication. */
  a->inode.use = PETSC_FALSE; /* Must disable: otherwise the MKL routines won't get used. */
  PetscCall(MatAssemblyEnd_SeqAIJ(A, mode));

  /* If the user has requested "eager" inspection, create the optimized MKL sparse handle (if needed; the function checks).
   * (The default is to do "lazy" inspection, deferring this until something like MatMult() is called.) */
  aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  if (aijmkl->eager_inspection) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));

  PetscFunctionReturn(0);
}

#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
PetscErrorCode MatMult_SeqAIJMKL(Mat A, Vec xx, Vec yy) {
  Mat_SeqAIJ        *a = (Mat_SeqAIJ *)A->data;
  const PetscScalar *x;
  PetscScalar       *y;
  const MatScalar   *aa;
  PetscInt           m     = A->rmap->n;
  PetscInt           n     = A->cmap->n;
  PetscScalar        alpha = 1.0;
  PetscScalar        beta  = 0.0;
  const PetscInt    *aj, *ai;
  char               matdescra[6];

  /* Variables not in MatMult_SeqAIJ. */
  char transa = 'n'; /* Used to indicate to MKL that we are not computing the transpose product. */

  PetscFunctionBegin;
  matdescra[0] = 'g'; /* Indicates to MKL that we using a general CSR matrix. */
  matdescra[3] = 'c'; /* Indicates to MKL that we use C-style (0-based) indexing. */
  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArray(yy, &y));
  aj = a->j; /* aj[k] gives column index for element aa[k]. */
  aa = a->a; /* Nonzero elements stored row-by-row. */
  ai = a->i; /* ai[k] is the position in aa and aj where row k starts. */

  /* Call MKL sparse BLAS routine to do the MatMult. */
  mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, y);

  PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArray(yy, &y));
  PetscFunctionReturn(0);
}
#endif

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
PetscErrorCode MatMult_SeqAIJMKL_SpMV2(Mat A, Vec xx, Vec yy) {
  Mat_SeqAIJ        *a      = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  const PetscScalar *x;
  PetscScalar       *y;
  PetscObjectState   state;

  PetscFunctionBegin;

  /* If there are no nonzero entries, zero yy and return immediately. */
  if (!a->nz) {
    PetscCall(VecGetArray(yy, &y));
    PetscCall(PetscArrayzero(y, A->rmap->n));
    PetscCall(VecRestoreArray(yy, &y));
    PetscFunctionReturn(0);
  }

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArray(yy, &y));

  /* In some cases, we get to this point without mkl_sparse_optimize() having been called, so we check and then call
   * it if needed. Eventually, when everything in PETSc is properly updating the matrix state, we should probably
   * take a "lazy" approach to creation/updating of the MKL matrix handle and plan to always do it here (when needed). */
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));

  /* Call MKL SpMV2 executor routine to do the MatMult. */
  PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, y);

  PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArray(yy, &y));
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
PetscErrorCode MatMultTranspose_SeqAIJMKL(Mat A, Vec xx, Vec yy) {
  Mat_SeqAIJ        *a = (Mat_SeqAIJ *)A->data;
  const PetscScalar *x;
  PetscScalar       *y;
  const MatScalar   *aa;
  PetscInt           m     = A->rmap->n;
  PetscInt           n     = A->cmap->n;
  PetscScalar        alpha = 1.0;
  PetscScalar        beta  = 0.0;
  const PetscInt    *aj, *ai;
  char               matdescra[6];

  /* Variables not in MatMultTranspose_SeqAIJ. */
  char transa = 't'; /* Used to indicate to MKL that we are computing the transpose product. */

  PetscFunctionBegin;
  matdescra[0] = 'g'; /* Indicates to MKL that we using a general CSR matrix. */
  matdescra[3] = 'c'; /* Indicates to MKL that we use C-style (0-based) indexing. */
  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArray(yy, &y));
  aj = a->j; /* aj[k] gives column index for element aa[k]. */
  aa = a->a; /* Nonzero elements stored row-by-row. */
  ai = a->i; /* ai[k] is the position in aa and aj where row k starts. */

  /* Call MKL sparse BLAS routine to do the MatMult. */
  mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, y);

  PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArray(yy, &y));
  PetscFunctionReturn(0);
}
#endif

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
PetscErrorCode MatMultTranspose_SeqAIJMKL_SpMV2(Mat A, Vec xx, Vec yy) {
  Mat_SeqAIJ        *a      = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  const PetscScalar *x;
  PetscScalar       *y;
  PetscObjectState   state;

  PetscFunctionBegin;

  /* If there are no nonzero entries, zero yy and return immediately. */
  if (!a->nz) {
    PetscCall(VecGetArray(yy, &y));
    PetscCall(PetscArrayzero(y, A->cmap->n));
    PetscCall(VecRestoreArray(yy, &y));
    PetscFunctionReturn(0);
  }

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArray(yy, &y));

  /* In some cases, we get to this point without mkl_sparse_optimize() having been called, so we check and then call
   * it if needed. Eventually, when everything in PETSc is properly updating the matrix state, we should probably
   * take a "lazy" approach to creation/updating of the MKL matrix handle and plan to always do it here (when needed). */
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));

  /* Call MKL SpMV2 executor routine to do the MatMultTranspose. */
  PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, y);

  PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArray(yy, &y));
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
PetscErrorCode MatMultAdd_SeqAIJMKL(Mat A, Vec xx, Vec yy, Vec zz) {
  Mat_SeqAIJ        *a = (Mat_SeqAIJ *)A->data;
  const PetscScalar *x;
  PetscScalar       *y, *z;
  const MatScalar   *aa;
  PetscInt           m = A->rmap->n;
  PetscInt           n = A->cmap->n;
  const PetscInt    *aj, *ai;
  PetscInt           i;

  /* Variables not in MatMultAdd_SeqAIJ. */
  char        transa = 'n'; /* Used to indicate to MKL that we are not computing the transpose product. */
  PetscScalar alpha  = 1.0;
  PetscScalar beta;
  char        matdescra[6];

  PetscFunctionBegin;
  matdescra[0] = 'g'; /* Indicates to MKL that we using a general CSR matrix. */
  matdescra[3] = 'c'; /* Indicates to MKL that we use C-style (0-based) indexing. */

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArrayPair(yy, zz, &y, &z));
  aj = a->j; /* aj[k] gives column index for element aa[k]. */
  aa = a->a; /* Nonzero elements stored row-by-row. */
  ai = a->i; /* ai[k] is the position in aa and aj where row k starts. */

  /* Call MKL sparse BLAS routine to do the MatMult. */
  if (zz == yy) {
    /* If zz and yy are the same vector, we can use MKL's mkl_xcsrmv(), which calculates y = alpha*A*x + beta*y. */
    beta = 1.0;
    mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, z);
  } else {
    /* zz and yy are different vectors, so call MKL's mkl_xcsrmv() with beta=0, then add the result to z.
     * MKL sparse BLAS does not have a MatMultAdd equivalent. */
    beta = 0.0;
    mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, z);
    for (i = 0; i < m; i++) z[i] += y[i];
  }

  PetscCall(PetscLogFlops(2.0 * a->nz));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
  PetscFunctionReturn(0);
}
#endif

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
PetscErrorCode MatMultAdd_SeqAIJMKL_SpMV2(Mat A, Vec xx, Vec yy, Vec zz) {
  Mat_SeqAIJ        *a      = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  const PetscScalar *x;
  PetscScalar       *y, *z;
  PetscInt           m = A->rmap->n;
  PetscInt           i;

  /* Variables not in MatMultAdd_SeqAIJ. */
  PetscObjectState state;

  PetscFunctionBegin;

  /* If there are no nonzero entries, set zz = yy and return immediately. */
  if (!a->nz) {
    PetscCall(VecGetArrayPair(yy, zz, &y, &z));
    PetscCall(PetscArraycpy(z, y, m));
    PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
    PetscFunctionReturn(0);
  }

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArrayPair(yy, zz, &y, &z));

  /* In some cases, we get to this point without mkl_sparse_optimize() having been called, so we check and then call
   * it if needed. Eventually, when everything in PETSc is properly updating the matrix state, we should probably
   * take a "lazy" approach to creation/updating of the MKL matrix handle and plan to always do it here (when needed). */
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));

  /* Call MKL sparse BLAS routine to do the MatMult. */
  if (zz == yy) {
    /* If zz and yy are the same vector, we can use mkl_sparse_x_mv, which calculates y = alpha*A*x + beta*y,
     * with alpha and beta both set to 1.0. */
    PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 1.0, z);
  } else {
    /* zz and yy are different vectors, so we call mkl_sparse_x_mv with alpha=1.0 and beta=0.0, and then
     * we add the contents of vector yy to the result; MKL sparse BLAS does not have a MatMultAdd equivalent. */
    PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, z);
    for (i = 0; i < m; i++) z[i] += y[i];
  }

  PetscCall(PetscLogFlops(2.0 * a->nz));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
PetscErrorCode MatMultTransposeAdd_SeqAIJMKL(Mat A, Vec xx, Vec yy, Vec zz) {
  Mat_SeqAIJ        *a = (Mat_SeqAIJ *)A->data;
  const PetscScalar *x;
  PetscScalar       *y, *z;
  const MatScalar   *aa;
  PetscInt           m = A->rmap->n;
  PetscInt           n = A->cmap->n;
  const PetscInt    *aj, *ai;
  PetscInt           i;

  /* Variables not in MatMultTransposeAdd_SeqAIJ. */
  char        transa = 't'; /* Used to indicate to MKL that we are computing the transpose product. */
  PetscScalar alpha  = 1.0;
  PetscScalar beta;
  char        matdescra[6];

  PetscFunctionBegin;
  matdescra[0] = 'g'; /* Indicates to MKL that we using a general CSR matrix. */
  matdescra[3] = 'c'; /* Indicates to MKL that we use C-style (0-based) indexing. */

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArrayPair(yy, zz, &y, &z));
  aj = a->j; /* aj[k] gives column index for element aa[k]. */
  aa = a->a; /* Nonzero elements stored row-by-row. */
  ai = a->i; /* ai[k] is the position in aa and aj where row k starts. */

  /* Call MKL sparse BLAS routine to do the MatMult. */
  if (zz == yy) {
    /* If zz and yy are the same vector, we can use MKL's mkl_xcsrmv(), which calculates y = alpha*A*x + beta*y. */
    beta = 1.0;
    mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, z);
  } else {
    /* zz and yy are different vectors, so call MKL's mkl_xcsrmv() with beta=0, then add the result to z.
     * MKL sparse BLAS does not have a MatMultAdd equivalent. */
    beta = 0.0;
    mkl_xcsrmv(&transa, &m, &n, &alpha, matdescra, aa, aj, ai, ai + 1, x, &beta, z);
    for (i = 0; i < n; i++) z[i] += y[i];
  }

  PetscCall(PetscLogFlops(2.0 * a->nz));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
  PetscFunctionReturn(0);
}
#endif

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
PetscErrorCode MatMultTransposeAdd_SeqAIJMKL_SpMV2(Mat A, Vec xx, Vec yy, Vec zz) {
  Mat_SeqAIJ        *a      = (Mat_SeqAIJ *)A->data;
  Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
  const PetscScalar *x;
  PetscScalar       *y, *z;
  PetscInt           n = A->cmap->n;
  PetscInt           i;
  PetscObjectState   state;

  /* Variables not in MatMultTransposeAdd_SeqAIJ. */

  PetscFunctionBegin;

  /* If there are no nonzero entries, set zz = yy and return immediately. */
  if (!a->nz) {
    PetscCall(VecGetArrayPair(yy, zz, &y, &z));
    PetscCall(PetscArraycpy(z, y, n));
    PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
    PetscFunctionReturn(0);
  }

  PetscCall(VecGetArrayRead(xx, &x));
  PetscCall(VecGetArrayPair(yy, zz, &y, &z));

  /* In some cases, we get to this point without mkl_sparse_optimize() having been called, so we check and then call
   * it if needed. Eventually, when everything in PETSc is properly updating the matrix state, we should probably
   * take a "lazy" approach to creation/updating of the MKL matrix handle and plan to always do it here (when needed). */
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!aijmkl->sparse_optimized || aijmkl->state != state) MatSeqAIJMKL_create_mkl_handle(A);

  /* Call MKL sparse BLAS routine to do the MatMult. */
  if (zz == yy) {
    /* If zz and yy are the same vector, we can use mkl_sparse_x_mv, which calculates y = alpha*A*x + beta*y,
     * with alpha and beta both set to 1.0. */
    PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 1.0, z);
  } else {
    /* zz and yy are different vectors, so we call mkl_sparse_x_mv with alpha=1.0 and beta=0.0, and then
     * we add the contents of vector yy to the result; MKL sparse BLAS does not have a MatMultAdd equivalent. */
    PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, z);
    for (i = 0; i < n; i++) z[i] += y[i];
  }

  PetscCall(PetscLogFlops(2.0 * a->nz));
  PetscCall(VecRestoreArrayRead(xx, &x));
  PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

/* -------------------------- MatProduct code -------------------------- */
#if defined(PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE)
static PetscErrorCode MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL_Private(Mat A, const sparse_operation_t transA, Mat B, const sparse_operation_t transB, Mat C) {
  Mat_SeqAIJMKL      *a = (Mat_SeqAIJMKL *)A->spptr, *b = (Mat_SeqAIJMKL *)B->spptr;
  sparse_matrix_t     csrA, csrB, csrC;
  PetscInt            nrows, ncols;
  struct matrix_descr descr_type_gen;
  PetscObjectState    state;

  PetscFunctionBegin;
  /* Determine the number of rows and columns that the result matrix C will have. We have to do this ourselves because MKL does
   * not handle sparse matrices with zero rows or columns. */
  if (transA == SPARSE_OPERATION_NON_TRANSPOSE) nrows = A->rmap->N;
  else nrows = A->cmap->N;
  if (transB == SPARSE_OPERATION_NON_TRANSPOSE) ncols = B->cmap->N;
  else ncols = B->rmap->N;

  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!a->sparse_optimized || a->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
  PetscCall(PetscObjectStateGet((PetscObject)B, &state));
  if (!b->sparse_optimized || b->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(B));
  csrA                = a->csrA;
  csrB                = b->csrA;
  descr_type_gen.type = SPARSE_MATRIX_TYPE_GENERAL;

  if (csrA && csrB) {
    PetscCallExternal(mkl_sparse_sp2m, transA, descr_type_gen, csrA, transB, descr_type_gen, csrB, SPARSE_STAGE_FULL_MULT_NO_VAL, &csrC);
  } else {
    csrC = PETSC_NULL;
  }

  PetscCall(MatSeqAIJMKL_setup_structure_from_mkl_handle(PETSC_COMM_SELF, csrC, nrows, ncols, C));

  PetscFunctionReturn(0);
}

PetscErrorCode MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL_Private(Mat A, const sparse_operation_t transA, Mat B, const sparse_operation_t transB, Mat C) {
  Mat_SeqAIJMKL      *a = (Mat_SeqAIJMKL *)A->spptr, *b = (Mat_SeqAIJMKL *)B->spptr, *c = (Mat_SeqAIJMKL *)C->spptr;
  sparse_matrix_t     csrA, csrB, csrC;
  struct matrix_descr descr_type_gen;
  PetscObjectState    state;

  PetscFunctionBegin;
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!a->sparse_optimized || a->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
  PetscCall(PetscObjectStateGet((PetscObject)B, &state));
  if (!b->sparse_optimized || b->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(B));
  csrA                = a->csrA;
  csrB                = b->csrA;
  csrC                = c->csrA;
  descr_type_gen.type = SPARSE_MATRIX_TYPE_GENERAL;

  if (csrA && csrB) {
    PetscCallExternal(mkl_sparse_sp2m, transA, descr_type_gen, csrA, transB, descr_type_gen, csrB, SPARSE_STAGE_FINALIZE_MULT, &csrC);
  } else {
    csrC = PETSC_NULL;
  }

  /* Have to update the PETSc AIJ representation for matrix C from contents of MKL handle. */
  PetscCall(MatSeqAIJMKL_update_from_mkl_handle(C));

  PetscFunctionReturn(0);
}

PetscErrorCode MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, PetscReal fill, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_NON_TRANSPOSE, B, SPARSE_OPERATION_NON_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

PetscErrorCode MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_NON_TRANSPOSE, B, SPARSE_OPERATION_NON_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

PetscErrorCode MatTransposeMatMultNumeric_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_TRANSPOSE, B, SPARSE_OPERATION_NON_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, PetscReal fill, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_TRANSPOSE, B, SPARSE_OPERATION_NON_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, PetscReal fill, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_NON_TRANSPOSE, B, SPARSE_OPERATION_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

PetscErrorCode MatMatTransposeMultNumeric_SeqAIJMKL_SeqAIJMKL(Mat A, Mat B, Mat C) {
  PetscFunctionBegin;
  PetscCall(MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL_Private(A, SPARSE_OPERATION_NON_TRANSPOSE, B, SPARSE_OPERATION_TRANSPOSE, C));
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductNumeric_AtB_SeqAIJMKL_SeqAIJMKL(Mat C) {
  Mat_Product *product = C->product;
  Mat          A = product->A, B = product->B;

  PetscFunctionBegin;
  PetscCall(MatTransposeMatMultNumeric_SeqAIJMKL_SeqAIJMKL(A, B, C));
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSymbolic_AtB_SeqAIJMKL_SeqAIJMKL(Mat C) {
  Mat_Product *product = C->product;
  Mat          A = product->A, B = product->B;
  PetscReal    fill = product->fill;

  PetscFunctionBegin;
  PetscCall(MatTransposeMatMultSymbolic_SeqAIJMKL_SeqAIJMKL(A, B, fill, C));
  C->ops->productnumeric = MatProductNumeric_AtB_SeqAIJMKL_SeqAIJMKL;
  PetscFunctionReturn(0);
}

PetscErrorCode MatPtAPNumeric_SeqAIJMKL_SeqAIJMKL_SymmetricReal(Mat A, Mat P, Mat C) {
  Mat                 Ct;
  Vec                 zeros;
  Mat_SeqAIJMKL      *a = (Mat_SeqAIJMKL *)A->spptr, *p = (Mat_SeqAIJMKL *)P->spptr, *c = (Mat_SeqAIJMKL *)C->spptr;
  sparse_matrix_t     csrA, csrP, csrC;
  PetscBool           set, flag;
  struct matrix_descr descr_type_sym;
  PetscObjectState    state;

  PetscFunctionBegin;
  PetscCall(MatIsSymmetricKnown(A, &set, &flag));
  PetscCheck(set && flag, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MatPtAPNumeric_SeqAIJMKL_SeqAIJMKL_SymmetricReal() called on matrix A not marked as symmetric");

  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!a->sparse_optimized || a->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
  PetscCall(PetscObjectStateGet((PetscObject)P, &state));
  if (!p->sparse_optimized || p->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(P));
  csrA                = a->csrA;
  csrP                = p->csrA;
  csrC                = c->csrA;
  descr_type_sym.type = SPARSE_MATRIX_TYPE_SYMMETRIC;
  descr_type_sym.mode = SPARSE_FILL_MODE_UPPER;
  descr_type_sym.diag = SPARSE_DIAG_NON_UNIT;

  /* Note that the call below won't work for complex matrices. (We protect this when pointers are assigned in MatConvert.) */
  PetscCallExternal(mkl_sparse_sypr, SPARSE_OPERATION_TRANSPOSE, csrP, csrA, descr_type_sym, &csrC, SPARSE_STAGE_FINALIZE_MULT);

  /* Update the PETSc AIJ representation for matrix C from contents of MKL handle.
   * This is more complicated than it should be: it turns out that, though mkl_sparse_sypr() will accept a full AIJ/CSR matrix,
   * the output matrix only contains the upper or lower triangle (we arbitrarily have chosen upper) of the symmetric matrix.
   * We have to fill in the missing portion, which we currently do below by forming the transpose and performing at MatAXPY
   * operation. This may kill any performance benefit of using the optimized mkl_sparse_sypr() routine. Performance might
   * improve if we come up with a more efficient way to do this, or we can convince the MKL team to provide an option to output
   * the full matrix. */
  PetscCall(MatSeqAIJMKL_update_from_mkl_handle(C));
  PetscCall(MatTranspose(C, MAT_INITIAL_MATRIX, &Ct));
  PetscCall(MatCreateVecs(C, &zeros, NULL));
  PetscCall(VecSetFromOptions(zeros));
  PetscCall(VecZeroEntries(zeros));
  PetscCall(MatDiagonalSet(Ct, zeros, INSERT_VALUES));
  PetscCall(MatAXPY(C, 1.0, Ct, DIFFERENT_NONZERO_PATTERN));
  /* Note: The MatAXPY() call destroys the MatProduct, so we must recreate it. */
  PetscCall(MatProductCreateWithMat(A, P, PETSC_NULL, C));
  PetscCall(MatProductSetType(C, MATPRODUCT_PtAP));
  PetscCall(MatSeqAIJMKL_create_mkl_handle(C));
  PetscCall(VecDestroy(&zeros));
  PetscCall(MatDestroy(&Ct));
  PetscFunctionReturn(0);
}

PetscErrorCode MatProductSymbolic_PtAP_SeqAIJMKL_SeqAIJMKL_SymmetricReal(Mat C) {
  Mat_Product        *product = C->product;
  Mat                 A = product->A, P = product->B;
  Mat_SeqAIJMKL      *a = (Mat_SeqAIJMKL *)A->spptr, *p = (Mat_SeqAIJMKL *)P->spptr;
  sparse_matrix_t     csrA, csrP, csrC;
  struct matrix_descr descr_type_sym;
  PetscObjectState    state;

  PetscFunctionBegin;
  PetscCall(PetscObjectStateGet((PetscObject)A, &state));
  if (!a->sparse_optimized || a->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
  PetscCall(PetscObjectStateGet((PetscObject)P, &state));
  if (!p->sparse_optimized || p->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(P));
  csrA                = a->csrA;
  csrP                = p->csrA;
  descr_type_sym.type = SPARSE_MATRIX_TYPE_SYMMETRIC;
  descr_type_sym.mode = SPARSE_FILL_MODE_UPPER;
  descr_type_sym.diag = SPARSE_DIAG_NON_UNIT;

  /* Note that the call below won't work for complex matrices. (We protect this when pointers are assigned in MatConvert.) */
  if (csrP && csrA) {
    PetscCallExternal(mkl_sparse_sypr, SPARSE_OPERATION_TRANSPOSE, csrP, csrA, descr_type_sym, &csrC, SPARSE_STAGE_FULL_MULT_NO_VAL);
  } else {
    csrC = PETSC_NULL;
  }

  /* Update the I and J arrays of the PETSc AIJ representation for matrix C from contents of MKL handle.
   * Note that, because mkl_sparse_sypr() only computes one triangle of the symmetric matrix, this representation will only contain
   * the upper triangle of the symmetric matrix. We fix this in MatPtAPNumeric_SeqAIJMKL_SeqAIJMKL_SymmetricReal(). I believe that
   * leaving things in this incomplete state is OK because the numeric product should follow soon after, but am not certain if this
   * is guaranteed. */
  PetscCall(MatSeqAIJMKL_setup_structure_from_mkl_handle(PETSC_COMM_SELF, csrC, P->cmap->N, P->cmap->N, C));

  C->ops->productnumeric = MatProductNumeric_PtAP;
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_AB(Mat C) {
  PetscFunctionBegin;
  C->ops->productsymbolic = MatProductSymbolic_AB;
  C->ops->matmultsymbolic = MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL;
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_AtB(Mat C) {
  PetscFunctionBegin;
  C->ops->productsymbolic = MatProductSymbolic_AtB_SeqAIJMKL_SeqAIJMKL;
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_ABt(Mat C) {
  PetscFunctionBegin;
  C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ;
  C->ops->productsymbolic          = MatProductSymbolic_ABt;
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_PtAP(Mat C) {
  Mat_Product *product = C->product;
  Mat          A       = product->A;
  PetscBool    set, flag;

  PetscFunctionBegin;
  if (PetscDefined(USE_COMPLEX)) {
    /* By setting C->ops->productsymbolic to NULL, we ensure that MatProductSymbolic_Unsafe() will be used.
     * We do this in several other locations in this file. This works for the time being, but these
     * routines are considered unsafe and may be removed from the MatProduct code in the future.
     * TODO: Add proper MATSEQAIJMKL implementations */
    C->ops->productsymbolic = NULL;
  } else {
    /* AIJMKL only has an optimized routine for PtAP when A is symmetric and real. */
    PetscCall(MatIsSymmetricKnown(A, &set, &flag));
    if (set && flag) C->ops->productsymbolic = MatProductSymbolic_PtAP_SeqAIJMKL_SeqAIJMKL_SymmetricReal;
    else C->ops->productsymbolic = NULL; /* MatProductSymbolic_Unsafe() will be used. */
    /* Note that we don't set C->ops->productnumeric here, as this must happen in MatProductSymbolic_PtAP_XXX(),
     * depending on whether the algorithm for the general case vs. the real symmetric one is used. */
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_RARt(Mat C) {
  PetscFunctionBegin;
  C->ops->productsymbolic = NULL; /* MatProductSymbolic_Unsafe() will be used. */
  PetscFunctionReturn(0);
}

static PetscErrorCode MatProductSetFromOptions_SeqAIJMKL_ABC(Mat C) {
  PetscFunctionBegin;
  C->ops->productsymbolic = NULL; /* MatProductSymbolic_Unsafe() will be used. */
  PetscFunctionReturn(0);
}

PetscErrorCode MatProductSetFromOptions_SeqAIJMKL(Mat C) {
  Mat_Product *product = C->product;

  PetscFunctionBegin;
  switch (product->type) {
  case MATPRODUCT_AB: PetscCall(MatProductSetFromOptions_SeqAIJMKL_AB(C)); break;
  case MATPRODUCT_AtB: PetscCall(MatProductSetFromOptions_SeqAIJMKL_AtB(C)); break;
  case MATPRODUCT_ABt: PetscCall(MatProductSetFromOptions_SeqAIJMKL_ABt(C)); break;
  case MATPRODUCT_PtAP: PetscCall(MatProductSetFromOptions_SeqAIJMKL_PtAP(C)); break;
  case MATPRODUCT_RARt: PetscCall(MatProductSetFromOptions_SeqAIJMKL_RARt(C)); break;
  case MATPRODUCT_ABC: PetscCall(MatProductSetFromOptions_SeqAIJMKL_ABC(C)); break;
  default: break;
  }
  PetscFunctionReturn(0);
}
#endif /* PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE */
/* ------------------------ End MatProduct code ------------------------ */

/* MatConvert_SeqAIJ_SeqAIJMKL converts a SeqAIJ matrix into a
 * SeqAIJMKL matrix.  This routine is called by the MatCreate_SeqAIJMKL()
 * routine, but can also be used to convert an assembled SeqAIJ matrix
 * into a SeqAIJMKL one. */
PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJMKL(Mat A, MatType type, MatReuse reuse, Mat *newmat) {
  Mat            B = *newmat;
  Mat_SeqAIJMKL *aijmkl;
  PetscBool      set;
  PetscBool      sametype;

  PetscFunctionBegin;
  if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &B));

  PetscCall(PetscObjectTypeCompare((PetscObject)A, type, &sametype));
  if (sametype) PetscFunctionReturn(0);

  PetscCall(PetscNewLog(B, &aijmkl));
  B->spptr = (void *)aijmkl;

  /* Set function pointers for methods that we inherit from AIJ but override.
   * We also parse some command line options below, since those determine some of the methods we point to. */
  B->ops->duplicate   = MatDuplicate_SeqAIJMKL;
  B->ops->assemblyend = MatAssemblyEnd_SeqAIJMKL;
  B->ops->destroy     = MatDestroy_SeqAIJMKL;

  aijmkl->sparse_optimized = PETSC_FALSE;
#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  aijmkl->no_SpMV2 = PETSC_FALSE; /* Default to using the SpMV2 routines if our MKL supports them. */
#else
  aijmkl->no_SpMV2 = PETSC_TRUE;
#endif
  aijmkl->eager_inspection = PETSC_FALSE;

  /* Parse command line options. */
  PetscOptionsBegin(PetscObjectComm((PetscObject)A), ((PetscObject)A)->prefix, "AIJMKL Options", "Mat");
  PetscCall(PetscOptionsBool("-mat_aijmkl_no_spmv2", "Disable use of inspector-executor (SpMV 2) routines", "None", (PetscBool)aijmkl->no_SpMV2, (PetscBool *)&aijmkl->no_SpMV2, &set));
  PetscCall(PetscOptionsBool("-mat_aijmkl_eager_inspection", "Run inspection at matrix assembly time, instead of waiting until needed by an operation", "None", (PetscBool)aijmkl->eager_inspection, (PetscBool *)&aijmkl->eager_inspection, &set));
  PetscOptionsEnd();
#if !defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  if (!aijmkl->no_SpMV2) {
    PetscCall(PetscInfo(B, "User requested use of MKL SpMV2 routines, but MKL version does not support mkl_sparse_optimize();  defaulting to non-SpMV2 routines.\n"));
    aijmkl->no_SpMV2 = PETSC_TRUE;
  }
#endif

#if defined(PETSC_HAVE_MKL_SPARSE_OPTIMIZE)
  B->ops->mult             = MatMult_SeqAIJMKL_SpMV2;
  B->ops->multtranspose    = MatMultTranspose_SeqAIJMKL_SpMV2;
  B->ops->multadd          = MatMultAdd_SeqAIJMKL_SpMV2;
  B->ops->multtransposeadd = MatMultTransposeAdd_SeqAIJMKL_SpMV2;
#if defined(PETSC_HAVE_MKL_SPARSE_SP2M_FEATURE)
  B->ops->productsetfromoptions   = MatProductSetFromOptions_SeqAIJMKL;
  B->ops->matmultsymbolic         = MatMatMultSymbolic_SeqAIJMKL_SeqAIJMKL;
  B->ops->matmultnumeric          = MatMatMultNumeric_SeqAIJMKL_SeqAIJMKL;
  B->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_SeqAIJMKL_SeqAIJMKL;
  B->ops->transposematmultnumeric = MatTransposeMatMultNumeric_SeqAIJMKL_SeqAIJMKL;
#if !defined(PETSC_USE_COMPLEX)
  B->ops->ptapnumeric = MatPtAPNumeric_SeqAIJMKL_SeqAIJMKL_SymmetricReal;
#else
  B->ops->ptapnumeric = NULL;
#endif
#endif
#endif /* PETSC_HAVE_MKL_SPARSE_OPTIMIZE */

#if !defined(PETSC_MKL_SPBLAS_DEPRECATED)
  /* In MKL version 18, update 2, the old sparse BLAS interfaces were marked as deprecated. If "no_SpMV2" has been specified by the
   * user and the old SpBLAS interfaces are deprecated in our MKL version, we use the new _SpMV2 routines (set above), but do not
   * call mkl_sparse_optimize(), which results in the old numerical kernels (without the inspector-executor model) being used. For
   * versions in which the older interface has not been deprecated, we use the old interface. */
  if (aijmkl->no_SpMV2) {
    B->ops->mult             = MatMult_SeqAIJMKL;
    B->ops->multtranspose    = MatMultTranspose_SeqAIJMKL;
    B->ops->multadd          = MatMultAdd_SeqAIJMKL;
    B->ops->multtransposeadd = MatMultTransposeAdd_SeqAIJMKL;
  }
#endif

  PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaijmkl_seqaij_C", MatConvert_SeqAIJMKL_SeqAIJ));

  PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJMKL));
  *newmat = B;
  PetscFunctionReturn(0);
}

/*@C
   MatCreateSeqAIJMKL - Creates a sparse matrix of type `MATSEQAIJMKL`.
   This type inherits from `MATSEQAIJ` and is largely identical, but uses sparse BLAS
   routines from Intel MKL whenever possible.
   If the installed version of MKL supports the "SpMV2" sparse
   inspector-executor routines, then those are used by default.
   `MatMult()`, `MatMultAdd()`, `MatMultTranspose()`, `MatMultTransposeAdd()`, `MatMatMult()`, `MatTransposeMatMult()`, and `MatPtAP()`
   (for symmetric A) operations are currently supported.
   Note that MKL version 18, update 2 or later is required for `MatPtAP()`, `MatPtAPNumeric()` and `MatMatMultNumeric()`.

   Collective

   Input Parameters:
+  comm - MPI communicator, set to `PETSC_COMM_SELF`
.  m - number of rows
.  n - number of columns
.  nz - number of nonzeros per row (same for all rows)
-  nnz - array containing the number of nonzeros in the various rows
         (possibly different for each row) or NULL

   Output Parameter:
.  A - the matrix

   Options Database Keys:
+  -mat_aijmkl_no_spmv2 - disable use of the SpMV2 inspector-executor routines
-  -mat_aijmkl_eager_inspection - perform MKL "inspection" phase upon matrix assembly; default is to do "lazy" inspection, performing this step the first time the matrix is applied

   Note:
   If nnz is given then nz is ignored

   Level: intermediate

.seealso: `MatCreate()`, `MatCreateMPIAIJMKL()`, `MatSetValues()`
@*/
PetscErrorCode MatCreateSeqAIJMKL(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A) {
  PetscFunctionBegin;
  PetscCall(MatCreate(comm, A));
  PetscCall(MatSetSizes(*A, m, n, m, n));
  PetscCall(MatSetType(*A, MATSEQAIJMKL));
  PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz));
  PetscFunctionReturn(0);
}

PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJMKL(Mat A) {
  PetscFunctionBegin;
  PetscCall(MatSetType(A, MATSEQAIJ));
  PetscCall(MatConvert_SeqAIJ_SeqAIJMKL(A, MATSEQAIJMKL, MAT_INPLACE_MATRIX, &A));
  PetscFunctionReturn(0);
}