Contents
sbscsm - block sparse column format triangular solve
SUBROUTINE SBSCSM( TRANSA, MB, N, UNITD, DV, ALPHA, DESCRA,
* VAL, BINDX, BPNTRB, BPNTRE, LB,
* B, LDB, BETA, C, LDC, WORK, LWORK )
INTEGER TRANSA, MB, N, UNITD, DESCRA(5), LB,
* LDB, LDC, LWORK
INTEGER BINDX(BNNZ), BPNTRB(MB), BPNTRE(MB)
REAL ALPHA, BETA
REAL DV(MB*LB*LB), VAL(LB*LB*BNNZ), B(LDB,*), C(LDC,*), WORK(LWORK)
SUBROUTINE SBSCSM_64( TRANSA, MB, N, UNITD, DV, ALPHA, DESCRA,
* VAL, BINDX, BPNTRB, BPNTRE, LB,
* B, LDB, BETA, C, LDC, WORK, LWORK )
INTEGER*8 TRANSA, MB, N, UNITD, DESCRA(5), LB,
* LDB, LDC, LWORK
INTEGER*8 BINDX(BNNZ), BPNTRB(MB), BPNTRE(MB)
REAL ALPHA, BETA
REAL DV(MB*LB*LB), VAL(LB*LB*BNNZ), B(LDB,*), C(LDC,*), WORK(LWORK)
where: BNNZ = BPNTRE(MB)- BPNTRB(1)
F95 INTERFACE
SUBROUTINE BSCSM(TRANSA, MB, [N], UNITD, DV, ALPHA, DESCRA, VAL, BINDX,
* BPNTRB, BPNTRE, LB, B, [LDB], BETA, C, [LDC], [WORK], [LWORK])
INTEGER TRANSA, MB, N, UNITD, LB
INTEGER, DIMENSION(:) :: DESCRA, BINDX, BPNTRB, BPNTRE
REAL ALPHA, BETA
REAL, DIMENSION(:) :: VAL, DV
REAL, DIMENSION(:, :) :: B, C
SUBROUTINE BSCSM_64(TRANSA, MB, [N], UNITD, DV, ALPHA, DESCRA, VAL, BINDX,
* BPNTRB, BPNTRE, LB, B, [LDB], BETA, C, [LDC], [WORK], [LWORK])
INTEGER*8 TRANSA, MB, N, UNITD, LB
INTEGER*8, DIMENSION(:) :: DESCRA, BINDX, BPNTRB, BPNTRE
REAL ALPHA, BETA
REAL, DIMENSION(:) :: VAL, DV
REAL, DIMENSION(:, :) :: B, C
C INTERFACE
#include <sunperf.h>
void sbscsm(int transa, int mb, int n, int unitd,
float *dv, float alpha, int *descra, float *val,
int *bindx, int *bpntrb, int *bpntre, int lb, float *b,
int ldb, float beta, float* c, int ldc);
void sbscsm_64(long transa, long mb, long n, long unitd,
float *dv, float alpha, long *descra, float *val, long
*bindx, long *bpntrb, long *bpntre, long lb, float *b, long
ldb, float beta, float *c, long ldc);
sbscsm performs one of the matrix-matrix operations
C <- alpha op(A) B + beta C, C <-alpha D op(A) B + beta C,
C <- alpha op(A) D B + beta C,
where alpha and beta are scalars, C and B are mb*lb by n dense matrices,
D is a block diagonal matrix, A is a sparse mb*lb by mb*lb unit, or
non-unit, upper or lower triangular matrix represented in the block
sparse column format and op( A ) is one of
op( A ) = inv(A) or op( A ) = inv(A') or op( A ) =inv(conjg( A' ))
(inv denotes matrix inverse, ' indicates matrix transpose).
TRANSA(input) On entry, integer TRANSA specifies the form of op(A) to be
used in the sparse matrix inverse as follows:.
0 : operate with matrix
1 : operate with transpose matrix
2 : operate with the conjugate transpose of matrix.
2 is equivalent to 1 if matrix is real.
Unchanged on exit.
MB(input) On entry, MB specifies the number of block columns
in the matrix A. Unchanged on exit.
N(input) On entry, N specifies the number of columns
in the matrix C. Unchanged on exit.
UNITD(input) On entry, integer UNITD specifies the type of scaling:
1 : Identity matrix (argument DV[] is ignored)
2 : Scale on left (row scaling)
3 : Scale on right (column scaling)
Unchanged on exit.
DV(input) On entry, DV is an array of length MB*LB*LB consisting
of the elements of the diagonal blocks of the matrix D.
The size of each square block is LB-by-LB and each
block is stored in standard column-major form.
Unchanged on exit.
ALPHA(input) On entry, ALPHA specifies the scalar alpha.
Unchanged on exit.
DESCRA (input) Descriptor argument. Five element integer array:
DESCRA(1) matrix structure
0 : general
1 : symmetric (A=A')
2 : Hermitian (A= CONJG(A'))
3 : Triangular
4 : Skew(Anti)-Symmetric (A=-A')
5 : Diagonal
6 : Skew-Hermitian (A= -CONJG(A'))
Note: For the routine, DESCRA(1)=3 is only supported.
DESCRA(2) upper/lower triangular indicator
1 : lower
2 : upper
DESCRA(3) main diagonal type
0 : non-identity blocks on the main diagonal
1 : identity diagonal blocks
2 : diagonal blocks are dense matrices
DESCRA(4) Array base (NOT IMPLEMENTED)
0 : C/C++ compatible
1 : Fortran compatible
DESCRA(5) repeated indices? (NOT IMPLEMENTED)
0 : unknown
1 : no repeated indices
VAL(input) On entry, VAL is a scalar array of length LB*LB*BNNZ
consisting of the non-zero block entries stored
column-major within each dense block where
BNNZ = BPNTRE(MB)-BPNTRB(1). Unchanged on exit.
BINDX(input) On entry, BINDX is an integer array of length BNNZ consisting
of the block row indices of the block entries of A where
BNNZ = BPNTRE(MB)-BPNTRB(1). The block row indices MUST
be sorted in increasing order for each block column.
Unchanged on exit.
BPNTRB(input) On entry,BPNTRB is an integer array of length MB such
that BPNTRB(J)-BPNTRB(1)+1 points to location in BINDX
of the first block entry of the J-th block column
of A. Unchanged on exit.
BPNTRE(input) On entry, BPNTRE is an integer array of length MB such
that BPNTRE(J)-BPNTRB(1) points to location in BINDX
of the last block entry of the J-th block column
of A. Unchanged on exit.
LB (input) On entry, LB specifies the dimension of dense blocks
composing A. Unchanged on exit.
B (input) Array of DIMENSION ( LDB, N ).
On entry, the leading mb*lb by n part of the array B
must contain the matrix B. Unchanged on exit.
LDB (input) On entry, LDB specifies the first dimension of B as declared
in the calling (sub) program. Unchanged on exit.
BETA (input) On entry, BETA specifies the scalar beta. Unchanged on exit.
C(input/output) Array of DIMENSION ( LDC, N ).
On entry, the leading mb*lb by n part of the array C
must contain the matrix C. On exit, the array C is
overwritten.
LDC (input) On entry, LDC specifies the first dimension of C as declared
in the calling (sub) program. Unchanged on exit.
WORK(workspace) Scratch array of length LWORK.
On exit, if LWORK= -1, WORK(1) returns the optimum size
of LWORK.
LWORK (input) On entry, LWORK specifies the length of WORK array. LWORK
should be at least MB*LB.
For good performance, LWORK should generally be larger.
For optimum performance on multiple processors, LWORK
>=MB*LB*N_CPUS where N_CPUS is the maximum number of
processors available to the program.
If LWORK=0, the routine is to allocate workspace needed.
If LWORK = -1, then a workspace query is assumed; the
routine only calculates the optimum size of the WORK array,
returns this value as the first entry of the WORK array,
and no error message related to LWORK is issued by XERBLA.
Libsunperf SPARSE BLAS is parallelized with the help of OPENMP and it is
fully compatible with NIST FORTRAN Sparse Blas but the sources are different.
Libsunperf SPARSE BLAS is free of bugs found in NIST FORTRAN Sparse Blas.
Besides several new features and routines are implemented.
NIST FORTRAN Sparse Blas User's Guide available at:
http://math.nist.gov/mcsd/Staff/KRemington/fspblas/
Based on the standard proposed in
"Document for the Basic Linear Algebra Subprograms (BLAS)
Standard", University of Tennessee, Knoxville, Tennessee, 1996:
http://www.netlib.org/utk/papers/sparse.ps
NOTES/BUGS
1. No test for singularity or near-singularity is included
in this routine. Such tests must be performed before calling
this routine.
2. If DESCRA(3)=0 , the lower or upper triangular part of
each diagonal block is used by the routine depending on
DESCRA(2) .
3. If DESCRA(3)=1 , the diagonal blocks in the block sparse
colum representation of A don't need to be the identity
matrices because these block entries are not used by the
routine in this case.
4. If DESCRA(3)=2 , the diagonal blocks are considered as
dense matrices and the LU factorization with partial
pivoting is used by the routine. WORK(1)=0 on return if the
factorization for all diagonal blocks has been completed
successfully, otherwise WORK(1) = - i where i is the block
number for which the LU factorization could not be computed.
5. The routine is designed so that it checks the validity of
each sparse block entry given in the sparse blas
representation. Block entries with incorrect indices are not
used and no error message related to the entries is issued.
The feature also provides a possibility to use the sparse
matrix representation of a general matrix A for solving
triangular systems with the upper or lower block triangle of
A. But DESCRA(1) MUST be equal to 3 even in this case.
Assume that there is the sparse matrix representation a
general matrix A decomposed in the form
A = L + D + U
where L is the strictly block lower triangle of A, U is the
strictly block upper triangle of A, D is the block diagonal
matrix. Let's I denotes the identity matrix.
Then the correspondence between the first three values of
DESCRA and the result matrix for the sparse representation
of A is
DESCRA(1) DESCRA(2) DESCRA(3) RESULT
3 1 1 alpha*op(L+I)*B+beta*C
3 1 0 alpha*op(L+D)*B+beta*C
3 2 1 alpha*op(U+I)*B+beta*C
3 2 0 alpha*op(U+D)*B+beta*C
6. It is known that there exists another representation of
the block sparse column format (see for example Y.Saad,
"Iterative Methods for Sparse Linear Systems", WPS, 1996).
Its data structure consists of three array instead of the
four used in the current implementation. The main
difference is that only one array, IA, containing the
pointers to the beginning of each block column in the arrays
VAL and BINDX is used instead of two arrays BPNTRB and
BPNTRE. To use the routine with this kind of block sparse
column format the following calling sequence should be used
CALL SBSCSM( TRANSA, MB, N, UNITD, DV, ALPHA, DESCRA,
* VAL, BINDX, IA, IA(2), LB,
* B, LDB, BETA, C, LDC, WORK, LWORK )