xref: /petsc/src/mat/impls/baij/mpi/baijmkl/mpibaijmkl.c (revision 7efe37a1cedd385a2f501b843d47cdf14dfb49ea)
1 #include <../src/mat/impls/baij/mpi/mpibaij.h>
2 
3 PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBAIJMKL(Mat, MatType, MatReuse, Mat *);
4 
5 static PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJMKL(Mat B, PetscInt bs, PetscInt d_nz, const PetscInt *d_nnz, PetscInt o_nz, const PetscInt *o_nnz)
6 {
7   Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data;
8 
9   PetscFunctionBegin;
10   PetscCall(MatMPIBAIJSetPreallocation_MPIBAIJ(B, bs, d_nz, d_nnz, o_nz, o_nnz));
11   PetscCall(MatConvert_SeqBAIJ_SeqBAIJMKL(b->A, MATSEQBAIJMKL, MAT_INPLACE_MATRIX, &b->A));
12   PetscCall(MatConvert_SeqBAIJ_SeqBAIJMKL(b->B, MATSEQBAIJMKL, MAT_INPLACE_MATRIX, &b->B));
13   PetscFunctionReturn(0);
14 }
15 
16 static PetscErrorCode MatConvert_MPIBAIJ_MPIBAIJMKL(Mat A, MatType type, MatReuse reuse, Mat *newmat)
17 {
18   Mat B = *newmat;
19 
20   PetscFunctionBegin;
21   if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &B));
22 
23   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIBAIJMKL));
24   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIBAIJSetPreallocation_C", MatMPIBAIJSetPreallocation_MPIBAIJMKL));
25   *newmat = B;
26   PetscFunctionReturn(0);
27 }
28 
29 /*@C
30    MatCreateBAIJMKL - Creates a sparse parallel matrix in `MATBAIJMKL` format
31    (block compressed row).
32    This type inherits from `MATBAIJ` and is largely identical, but uses sparse BLAS
33    routines from Intel MKL whenever possible.
34    `MatMult()`, `MatMultAdd()`, `MatMultTranspose()`, and `MatMultTransposeAdd()`
35    operations are currently supported.
36    If the installed version of MKL supports the "SpMV2" sparse
37    inspector-executor routines, then those are used by default.
38    Default PETSc kernels are used otherwise.
39    For good matrix assembly performance the user should preallocate the matrix
40    storage by setting the parameters d_nz (or d_nnz) and o_nz (or o_nnz).
41    By setting these parameters accurately, performance can be increased by more
42    than a factor of 50.
43 
44    Collective
45 
46    Input Parameters:
47 +  comm - MPI communicator
48 .  bs   - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
49           blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
50 .  m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
51            This value should be the same as the local size used in creating the
52            y vector for the matrix-vector product y = Ax.
53 .  n - number of local columns (or `PETSC_DECIDE` to have calculated if N is given)
54            This value should be the same as the local size used in creating the
55            x vector for the matrix-vector product y = Ax.
56 .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
57 .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
58 .  d_nz  - number of nonzero blocks per block row in diagonal portion of local
59            submatrix  (same for all local rows)
60 .  d_nnz - array containing the number of nonzero blocks in the various block rows
61            of the in diagonal portion of the local (possibly different for each block
62            row) or NULL.  If you plan to factor the matrix you must leave room for the diagonal entry
63            and set it even if it is zero.
64 .  o_nz  - number of nonzero blocks per block row in the off-diagonal portion of local
65            submatrix (same for all local rows).
66 -  o_nnz - array containing the number of nonzero blocks in the various block rows of the
67            off-diagonal portion of the local submatrix (possibly different for
68            each block row) or NULL.
69 
70    Output Parameter:
71 .  A - the matrix
72 
73    Options Database Keys:
74 +   -mat_block_size - size of the blocks to use
75 -   -mat_use_hash_table <fact> - set hash table factor
76 
77    It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
78    MatXXXXSetPreallocation() paradigm instead of this routine directly.
79    [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
80 
81    Notes:
82    If the *_nnz parameter is given then the *_nz parameter is ignored
83 
84    A nonzero block is any block that as 1 or more nonzeros in it
85 
86    The user MUST specify either the local or global matrix dimensions
87    (possibly both).
88 
89    If `PETSC_DECIDE` or  `PETSC_DETERMINE` is used for a particular argument on one processor
90    than it must be used on all processors that share the object for that argument.
91 
92    Storage Information:
93    For a square global matrix we define each processor's diagonal portion
94    to be its local rows and the corresponding columns (a square submatrix);
95    each processor's off-diagonal portion encompasses the remainder of the
96    local matrix (a rectangular submatrix).
97 
98    The user can specify preallocated storage for the diagonal part of
99    the local submatrix with either d_nz or d_nnz (not both).  Set
100    d_nz = `PETSC_DEFAULT` and d_nnz = NULL for PETSc to control dynamic
101    memory allocation.  Likewise, specify preallocated storage for the
102    off-diagonal part of the local submatrix with o_nz or o_nnz (not both).
103 
104    Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
105    the figure below we depict these three local rows and all columns (0-11).
106 
107 .vb
108            0 1 2 3 4 5 6 7 8 9 10 11
109           --------------------------
110    row 3  |o o o d d d o o o o  o  o
111    row 4  |o o o d d d o o o o  o  o
112    row 5  |o o o d d d o o o o  o  o
113           --------------------------
114 .ve
115 
116    Thus, any entries in the d locations are stored in the d (diagonal)
117    submatrix, and any entries in the o locations are stored in the
118    o (off-diagonal) submatrix.  Note that the d and the o submatrices are
119    stored simply in the `MATSEQBAIJMKL` format for compressed row storage.
120 
121    Now d_nz should indicate the number of block nonzeros per row in the d matrix,
122    and o_nz should indicate the number of block nonzeros per row in the o matrix.
123    In general, for PDE problems in which most nonzeros are near the diagonal,
124    one expects d_nz >> o_nz.   For large problems you MUST preallocate memory
125    or you will get TERRIBLE performance; see the users' manual chapter on
126    matrices.
127 
128    Level: intermediate
129 
130 .seealso: `MATBAIJMKL`, `MATBAIJ`, `MatCreate()`, `MatCreateSeqBAIJMKL()`, `MatSetValues()`, `MatCreateBAIJMKL()`, `MatMPIBAIJSetPreallocation()`, `MatMPIBAIJSetPreallocationCSR()`
131 @*/
132 
133 PetscErrorCode MatCreateBAIJMKL(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[], Mat *A)
134 {
135   PetscMPIInt size;
136 
137   PetscFunctionBegin;
138   PetscCall(MatCreate(comm, A));
139   PetscCall(MatSetSizes(*A, m, n, M, N));
140   PetscCallMPI(MPI_Comm_size(comm, &size));
141   if (size > 1) {
142     PetscCall(MatSetType(*A, MATMPIBAIJMKL));
143     PetscCall(MatMPIBAIJSetPreallocation(*A, bs, d_nz, d_nnz, o_nz, o_nnz));
144   } else {
145     PetscCall(MatSetType(*A, MATSEQBAIJMKL));
146     PetscCall(MatSeqBAIJSetPreallocation(*A, bs, d_nz, d_nnz));
147   }
148   PetscFunctionReturn(0);
149 }
150 
151 PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJMKL(Mat A)
152 {
153   PetscFunctionBegin;
154   PetscCall(MatSetType(A, MATMPIBAIJ));
155   PetscCall(MatConvert_MPIBAIJ_MPIBAIJMKL(A, MATMPIBAIJMKL, MAT_INPLACE_MATRIX, &A));
156   PetscFunctionReturn(0);
157 }
158 
159 /*MC
160    MATBAIJMKL - MATBAIJMKL = "BAIJMKL" - A matrix type to be used for sparse matrices.
161 
162    This matrix type is identical to `MATSEQBAIJMKL` when constructed with a single process communicator,
163    and `MATMPIBAIJMKL` otherwise.  As a result, for single process communicators,
164   `MatSeqBAIJSetPreallocation()` is supported, and similarly `MatMPIBAIJSetPreallocation()` is supported
165   for communicators controlling multiple processes.  It is recommended that you call both of
166   the above preallocation routines for simplicity.
167 
168    Options Database Keys:
169 . -mat_type baijmkl - sets the matrix type to `MATBAIJMKL` during a call to `MatSetFromOptions()`
170 
171   Level: beginner
172 
173 .seealso: `MatCreateBAIJMKL()`, `MATSEQBAIJMKL`, `MATMPIBAIJMKL`
174 M*/
175