xref: /petsc/src/mat/impls/sell/mpi/mpicuda/mpisellcuda.cu (revision 2d1451d43b73a0495cd81c074cbc1e0206888947)

#include <petscconf.h>
#include <../src/mat/impls/sell/mpi/mpisell.h> /*I "petscmat.h" I*/

PetscErrorCode MatMPISELLSetPreallocation_MPISELLCUDA(Mat B, PetscInt d_rlenmax, const PetscInt d_rlen[], PetscInt o_rlenmax, const PetscInt o_rlen[])
{
  Mat_MPISELL *b = (Mat_MPISELL *)B->data;

  PetscFunctionBegin;
  PetscCall(PetscLayoutSetUp(B->rmap));
  PetscCall(PetscLayoutSetUp(B->cmap));

  if (!B->preallocated) {
    /* Explicitly create 2 MATSEQSELLCUDA matrices. */
    PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
    PetscCall(MatBindToCPU(b->A, B->boundtocpu));
    PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
    PetscCall(MatSetType(b->A, MATSEQSELLCUDA));
    PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
    PetscCall(MatBindToCPU(b->B, B->boundtocpu));
    PetscCall(MatSetSizes(b->B, B->rmap->n, B->cmap->N, B->rmap->n, B->cmap->N));
    PetscCall(MatSetType(b->B, MATSEQSELLCUDA));
  }
  PetscCall(MatSeqSELLSetPreallocation(b->A, d_rlenmax, d_rlen));
  PetscCall(MatSeqSELLSetPreallocation(b->B, o_rlenmax, o_rlen));
  B->preallocated  = PETSC_TRUE;
  B->was_assembled = PETSC_FALSE;
  B->assembled     = PETSC_FALSE;
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatMult_MPISELLCUDA(Mat A, Vec xx, Vec yy)
{
  Mat_MPISELL *a = (Mat_MPISELL *)A->data;
  PetscInt     nt;

  PetscFunctionBegin;
  PetscCall(VecGetLocalSize(xx, &nt));
  PetscCheck(nt == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A (%" PetscInt_FMT ") and xx (%" PetscInt_FMT ")", A->cmap->n, nt);
  PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
  PetscCall((*a->A->ops->mult)(a->A, xx, yy));
  PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
  PetscCall((*a->B->ops->multadd)(a->B, a->lvec, yy, yy));
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatMultAdd_MPISELLCUDA(Mat A, Vec xx, Vec yy, Vec zz)
{
  Mat_MPISELL *a = (Mat_MPISELL *)A->data;
  PetscInt     nt;

  PetscFunctionBegin;
  PetscCall(VecGetLocalSize(xx, &nt));
  PetscCheck(nt == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A (%" PetscInt_FMT ") and xx (%" PetscInt_FMT ")", A->cmap->n, nt);
  PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
  PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
  PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
  PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatMultTranspose_MPISELLCUDA(Mat A, Vec xx, Vec yy)
{
  Mat_MPISELL *a = (Mat_MPISELL *)A->data;
  PetscInt     nt;

  PetscFunctionBegin;
  PetscCall(VecGetLocalSize(xx, &nt));
  PetscCheck(nt == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A (%" PetscInt_FMT ") and xx (%" PetscInt_FMT ")", A->rmap->n, nt);
  PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
  PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
  PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
  PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatSetFromOptions_MPISELLCUDA(PetscOptionItems *PetscOptionsObject, Mat A)
{
  PetscFunctionBegin;
  PetscOptionsHeadBegin(PetscOptionsObject, "MPISELLCUDA options");
  if (A->factortype == MAT_FACTOR_NONE) { }
  PetscOptionsHeadEnd();
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatAssemblyEnd_MPISELLCUDA(Mat A, MatAssemblyType mode)
{
  Mat_MPISELL *mpisell;

  PetscFunctionBegin;
  mpisell = (Mat_MPISELL *)A->data;
  PetscCall(MatAssemblyEnd_MPISELL(A, mode));
  if (!A->was_assembled && mode == MAT_FINAL_ASSEMBLY) { PetscCall(VecSetType(mpisell->lvec, VECSEQCUDA)); }
  PetscFunctionReturn(PETSC_SUCCESS);
}

PetscErrorCode MatDestroy_MPISELLCUDA(Mat A)
{
  PetscFunctionBegin;
  PetscCall(MatDestroy_MPISELL(A));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpisellcuda_mpiaij_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPISELLSetPreallocation_C", NULL));
  PetscFunctionReturn(PETSC_SUCCESS);
}

PETSC_EXTERN PetscErrorCode MatCreate_MPISELLCUDA(Mat A)
{
  PetscFunctionBegin;
  PetscCall(MatCreate_MPISELL(A));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPISELLSetPreallocation_C", MatMPISELLSetPreallocation_MPISELLCUDA));
  PetscCall(PetscFree(A->defaultvectype));
  PetscCall(PetscStrallocpy(VECCUDA, &A->defaultvectype));

  A->ops->assemblyend   = MatAssemblyEnd_MPISELLCUDA;
  A->ops->mult          = MatMult_MPISELLCUDA;
  A->ops->multadd       = MatMultAdd_MPISELLCUDA;
  A->ops->multtranspose = MatMultTranspose_MPISELLCUDA;
  A->ops->destroy       = MatDestroy_MPISELLCUDA;

  PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATMPISELLCUDA));
  PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpisellcuda_mpiaij_C", MatConvert_MPISELL_MPIAIJ));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
   MatCreateSELLCUDA - Creates a sparse matrix in SELL format.
   This matrix will ultimately pushed down to NVIDIA GPUs.

   Collective

   Input Parameters:
+  comm - MPI communicator, set to `PETSC_COMM_SELF`
.  m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
           This value should be the same as the local size used in creating the
           y vector for the matrix-vector product y = Ax.
.  n - This value should be the same as the local size used in creating the
       x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
       calculated if `N` is given) For square matrices `n` is almost always `m`.
.  M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
.  N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
.  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
           (same value is used for all local rows)
.  d_nnz - array containing the number of nonzeros in the various rows of the
           DIAGONAL portion of the local submatrix (possibly different for each row)
           or `NULL`, if `d_nz` is used to specify the nonzero structure.
           The size of this array is equal to the number of local rows, i.e `m`.
           For matrices you plan to factor you must leave room for the diagonal entry and
           put in the entry even if it is zero.
.  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
           submatrix (same value is used for all local rows).
-  o_nnz - array containing the number of nonzeros in the various rows of the
           OFF-DIAGONAL portion of the local submatrix (possibly different for
           each row) or `NULL`, if `o_nz` is used to specify the nonzero
           structure. The size of this array is equal to the number
           of local rows, i.e `m`.

   Output Parameter:
.  A - the matrix

   Level: intermediate

   Notes:
   If `nnz` is given then `nz` is ignored

   Specify the preallocated storage with either `nz` or `nnz` (not both).
   Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
   allocation.

.seealso: [](chapter_matrices), `Mat`, `MatCreate()`, `MatCreateSELL()`, `MatSetValues()`, `MATMPISELLCUDA`, `MATSELLCUDA`
@*/
PetscErrorCode MatCreateSELLCUDA(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[], Mat *A)
{
  PetscMPIInt size;

  PetscFunctionBegin;
  PetscCall(MatCreate(comm, A));
  PetscCall(MatSetSizes(*A, m, n, M, N));
  PetscCallMPI(MPI_Comm_size(comm, &size));
  if (size > 1) {
    PetscCall(MatSetType(*A, MATMPISELLCUDA));
    PetscCall(MatMPISELLSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
  } else {
    PetscCall(MatSetType(*A, MATSEQSELLCUDA));
    PetscCall(MatSeqSELLSetPreallocation(*A, d_nz, d_nnz));
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*MC
   MATSELLCUDA - "sellcuda" = "mpisellcuda" - A matrix type to be used for sparse matrices.

   Sliced ELLPACK matrix type whose data resides on NVIDIA GPUs.

   This matrix type is identical to `MATSEQSELLCUDA` when constructed with a single process communicator,
   and `MATMPISELLCUDA` otherwise.  As a result, for single process communicators,
   `MatSeqSELLSetPreallocation()` is supported, and similarly `MatMPISELLSetPreallocation()` is supported
   for communicators controlling multiple processes.  It is recommended that you call both of
   the above preallocation routines for simplicity.

   Options Database Key:
.  -mat_type mpisellcuda - sets the matrix type to `MATMPISELLCUDA` during a call to MatSetFromOptions()

  Level: beginner

 .seealso: `MatCreateSELLCUDA()`, `MATSEQSELLCUDA`, `MatCreateSeqSELLCUDA()`, `MatCUDAFormatOperation()`
M*/