*DECK DSLUCS
SUBROUTINE DSLUCS (N, B, X, NELT, IA, JA, A, ISYM, ITOL, TOL,
+ ITMAX, ITER, ERR, IERR, IUNIT, RWORK, LENW, IWORK, LENIW)
C***BEGIN PROLOGUE DSLUCS
C***PURPOSE Incomplete LU BiConjugate Gradient Squared Ax=b Solver.
C Routine to solve a linear system Ax = b using the
C BiConjugate Gradient Squared method with Incomplete LU
C decomposition preconditioning.
C***LIBRARY SLATEC (SLAP)
C***CATEGORY D2A4, D2B4
C***TYPE DOUBLE PRECISION (SSLUCS-S, DSLUCS-D)
C***KEYWORDS ITERATIVE INCOMPLETE LU PRECONDITION,
C NON-SYMMETRIC LINEAR SYSTEM, SLAP, SPARSE
C***AUTHOR Greenbaum, Anne, (Courant Institute)
C Seager, Mark K., (LLNL)
C Lawrence Livermore National Laboratory
C PO BOX 808, L-60
C Livermore, CA 94550 (510) 423-3141
C seager@llnl.gov
C***DESCRIPTION
C
C *Usage:
C INTEGER N, NELT, IA(NELT), JA(NELT), ISYM, ITOL, ITMAX
C INTEGER ITER, IERR, IUNIT, LENW, IWORK(NL+NU+4*N+2), LENIW
C DOUBLE PRECISION B(N), X(N), A(NELT), TOL, ERR, RWORK(NL+NU+8*N)
C
C CALL DSLUCS(N, B, X, NELT, IA, JA, A, ISYM, ITOL, TOL,
C $ ITMAX, ITER, ERR, IERR, IUNIT, RWORK, LENW, IWORK, LENIW)
C
C *Arguments:
C N :IN Integer.
C Order of the Matrix.
C B :IN Double Precision B(N).
C Right-hand side vector.
C X :INOUT Double Precision X(N).
C On input X is your initial guess for solution vector.
C On output X is the final approximate solution.
C NELT :IN Integer.
C Number of Non-Zeros stored in A.
C IA :INOUT Integer IA(NELT).
C JA :INOUT Integer JA(NELT).
C A :INOUT Double Precision A(NELT).
C These arrays should hold the matrix A in either the SLAP
C Triad format or the SLAP Column format. See "Description",
C below. If the SLAP Triad format is chosen it is changed
C internally to the SLAP Column format.
C ISYM :IN Integer.
C Flag to indicate symmetric storage format.
C If ISYM=0, all non-zero entries of the matrix are stored.
C If ISYM=1, the matrix is symmetric, and only the upper
C or lower triangle of the matrix is stored.
C ITOL :IN Integer.
C Flag to indicate type of convergence criterion.
C If ITOL=1, iteration stops when the 2-norm of the residual
C divided by the 2-norm of the right-hand side is less than TOL.
C This routine must calculate the residual from R = A*X - B.
C This is unnatural and hence expensive for this type of iter-
C ative method. ITOL=2 is *STRONGLY* recommended.
C If ITOL=2, iteration stops when the 2-norm of M-inv times the
C residual divided by the 2-norm of M-inv times the right hand
C side is less than TOL, where M-inv time a vector is the pre-
C conditioning step. This is the *NATURAL* stopping for this
C iterative method and is *STRONGLY* recommended.
C TOL :INOUT Double Precision.
C Convergence criterion, as described above. (Reset if IERR=4.)
C ITMAX :IN Integer.
C Maximum number of iterations.
C ITER :OUT Integer.
C Number of iterations required to reach convergence, or
C ITMAX+1 if convergence criterion could not be achieved in
C ITMAX iterations.
C ERR :OUT Double Precision.
C Error estimate of error in final approximate solution, as
C defined by ITOL.
C IERR :OUT Integer.
C Return error flag.
C IERR = 0 => All went well.
C IERR = 1 => Insufficient space allocated for WORK or IWORK.
C IERR = 2 => Method failed to converge in ITMAX steps.
C IERR = 3 => Error in user input.
C Check input values of N, ITOL.
C IERR = 4 => User error tolerance set too tight.
C Reset to 500*D1MACH(3). Iteration proceeded.
C IERR = 5 => Breakdown of the method detected.
C (r0,r) approximately 0.
C IERR = 6 => Stagnation of the method detected.
C (r0,v) approximately 0.
C IERR = 7 => Incomplete factorization broke down and was
C fudged. Resulting preconditioning may be less
C than the best.
C IUNIT :IN Integer.
C Unit number on which to write the error at each iteration,
C if this is desired for monitoring convergence. If unit
C number is 0, no writing will occur.
C RWORK :WORK Double Precision RWORK(LENW).
C Double Precision array used for workspace. NL is the number
C of non-zeros in the lower triangle of the matrix (including
C the diagonal). NU is the number of non-zeros in the upper
C triangle of the matrix (including the diagonal).
C LENW :IN Integer.
C Length of the double precision workspace, RWORK.
C LENW >= NL+NU+8*N.
C IWORK :WORK Integer IWORK(LENIW).
C Integer array used for workspace. NL is the number of non-
C zeros in the lower triangle of the matrix (including the
C diagonal). NU is the number of non-zeros in the upper
C triangle of the matrix (including the diagonal).
C Upon return the following locations of IWORK hold information
C which may be of use to the user:
C IWORK(9) Amount of Integer workspace actually used.
C IWORK(10) Amount of Double Precision workspace actually used.
C LENIW :IN Integer.
C Length of the integer workspace, IWORK.
C LENIW >= NL+NU+4*N+12.
C
C *Description:
C This routine is simply a driver for the DCGSN routine. It
C calls the DSILUS routine to set up the preconditioning and
C then calls DCGSN with the appropriate MATVEC, MTTVEC and
C MSOLVE, MTSOLV routines.
C
C The Sparse Linear Algebra Package (SLAP) utilizes two matrix
C data structures: 1) the SLAP Triad format or 2) the SLAP
C Column format. The user can hand this routine either of the
C of these data structures and SLAP will figure out which on
C is being used and act accordingly.
C
C =================== S L A P Triad format ===================
C
C This routine requires that the matrix A be stored in the
C SLAP Triad format. In this format only the non-zeros are
C stored. They may appear in *ANY* order. The user supplies
C three arrays of length NELT, where NELT is the number of
C non-zeros in the matrix: (IA(NELT), JA(NELT), A(NELT)). For
C each non-zero the user puts the row and column index of that
C matrix element in the IA and JA arrays. The value of the
C non-zero matrix element is placed in the corresponding
C location of the A array. This is an extremely easy data
C structure to generate. On the other hand it is not too
C efficient on vector computers for the iterative solution of
C linear systems. Hence, SLAP changes this input data
C structure to the SLAP Column format for the iteration (but
C does not change it back).
C
C Here is an example of the SLAP Triad storage format for a
C 5x5 Matrix. Recall that the entries may appear in any order.
C
C 5x5 Matrix SLAP Triad format for 5x5 matrix on left.
C 1 2 3 4 5 6 7 8 9 10 11
C |11 12 0 0 15| A: 51 12 11 33 15 53 55 22 35 44 21
C |21 22 0 0 0| IA: 5 1 1 3 1 5 5 2 3 4 2
C | 0 0 33 0 35| JA: 1 2 1 3 5 3 5 2 5 4 1
C | 0 0 0 44 0|
C |51 0 53 0 55|
C
C =================== S L A P Column format ==================
C
C This routine requires that the matrix A be stored in the
C SLAP Column format. In this format the non-zeros are stored
C counting down columns (except for the diagonal entry, which
C must appear first in each "column") and are stored in the
C double precision array A. In other words, for each column
C in the matrix put the diagonal entry in A. Then put in the
C other non-zero elements going down the column (except the
C diagonal) in order. The IA array holds the row index for
C each non-zero. The JA array holds the offsets into the IA,
C A arrays for the beginning of each column. That is,
C IA(JA(ICOL)), A(JA(ICOL)) points to the beginning of the
C ICOL-th column in IA and A. IA(JA(ICOL+1)-1),
C A(JA(ICOL+1)-1) points to the end of the ICOL-th column.
C Note that we always have JA(N+1) = NELT+1, where N is the
C number of columns in the matrix and NELT is the number of
C non-zeros in the matrix.
C
C Here is an example of the SLAP Column storage format for a
C 5x5 Matrix (in the A and IA arrays '|' denotes the end of a
C column):
C
C 5x5 Matrix SLAP Column format for 5x5 matrix on left.
C 1 2 3 4 5 6 7 8 9 10 11
C |11 12 0 0 15| A: 11 21 51 | 22 12 | 33 53 | 44 | 55 15 35
C |21 22 0 0 0| IA: 1 2 5 | 2 1 | 3 5 | 4 | 5 1 3
C | 0 0 33 0 35| JA: 1 4 6 8 9 12
C | 0 0 0 44 0|
C |51 0 53 0 55|
C
C *Side Effects:
C The SLAP Triad format (IA, JA, A) is modified internally to
C be the SLAP Column format. See above.
C
C *Cautions:
C This routine will attempt to write to the Fortran logical output
C unit IUNIT, if IUNIT .ne. 0. Thus, the user must make sure that
C this logical unit is attached to a file or terminal before calling
C this routine with a non-zero value for IUNIT. This routine does
C not check for the validity of a non-zero IUNIT unit number.
C
C***SEE ALSO DCGS, DSDCGS
C***REFERENCES 1. P. Sonneveld, CGS, a fast Lanczos-type solver
C for nonsymmetric linear systems, Delft University
C of Technology Report 84-16, Department of Mathe-
C matics and Informatics, Delft, The Netherlands.
C 2. E. F. Kaasschieter, The solution of non-symmetric
C linear systems by biconjugate gradients or conjugate
C gradients squared, Delft University of Technology
C Report 86-21, Department of Mathematics and Informa-
C tics, Delft, The Netherlands.
C***ROUTINES CALLED DCGS, DCHKW, DS2Y, DSILUS, DSLUI, DSMV
C***REVISION HISTORY (YYMMDD)
C 890404 DATE WRITTEN
C 890404 Previous REVISION DATE
C 890915 Made changes requested at July 1989 CML Meeting. (MKS)
C 890921 Removed TeX from comments. (FNF)
C 890922 Numerous changes to prologue to make closer to SLATEC
C standard. (FNF)
C 890929 Numerous changes to reduce SP/DP differences. (FNF)
C 910411 Prologue converted to Version 4.0 format. (BAB)
C 920511 Added complete declaration section. (WRB)
C 920929 Corrected format of references. (FNF)
C 921113 Corrected C***CATEGORY line. (FNF)
C***END PROLOGUE DSLUCS
C .. Parameters ..
INTEGER LOCRB, LOCIB
PARAMETER (LOCRB=1, LOCIB=11)
C .. Scalar Arguments ..
DOUBLE PRECISION ERR, TOL
INTEGER IERR, ISYM, ITER, ITMAX, ITOL, IUNIT, LENIW, LENW, N, NELT
C .. Array Arguments ..
DOUBLE PRECISION A(NELT), B(N), RWORK(LENW), X(N)
INTEGER IA(NELT), IWORK(LENIW), JA(NELT)
C .. Local Scalars ..
INTEGER ICOL, J, JBGN, JEND, LOCDIN, LOCIL, LOCIU, LOCIW, LOCJL,
+ LOCJU, LOCL, LOCNC, LOCNR, LOCP, LOCQ, LOCR, LOCR0, LOCU,
+ LOCUU, LOCV1, LOCV2, LOCW, NL, NU
C .. External Subroutines ..
EXTERNAL DCGS, DCHKW, DS2Y, DSILUS, DSLUI, DSMV
C***FIRST EXECUTABLE STATEMENT DSLUCS
C
IERR = 0
IF( N.LT.1 .OR. NELT.LT.1 ) THEN
IERR = 3
RETURN
ENDIF
C
C Change the SLAP input matrix IA, JA, A to SLAP-Column format.
CALL DS2Y( N, NELT, IA, JA, A, ISYM )
C
C Count number of Non-Zero elements preconditioner ILU matrix.
C Then set up the work arrays.
NL = 0
NU = 0
DO 20 ICOL = 1, N
C Don't count diagonal.
JBGN = JA(ICOL)+1
JEND = JA(ICOL+1)-1
IF( JBGN.LE.JEND ) THEN
CVD$ NOVECTOR
DO 10 J = JBGN, JEND
IF( IA(J).GT.ICOL ) THEN
NL = NL + 1
IF( ISYM.NE.0 ) NU = NU + 1
ELSE
NU = NU + 1
ENDIF
10 CONTINUE
ENDIF
20 CONTINUE
C
LOCIL = LOCIB
LOCJL = LOCIL + N+1
LOCIU = LOCJL + NL
LOCJU = LOCIU + NU
LOCNR = LOCJU + N+1
LOCNC = LOCNR + N
LOCIW = LOCNC + N
C
LOCL = LOCRB
LOCDIN = LOCL + NL
LOCUU = LOCDIN + N
LOCR = LOCUU + NU
LOCR0 = LOCR + N
LOCP = LOCR0 + N
LOCQ = LOCP + N
LOCU = LOCQ + N
LOCV1 = LOCU + N
LOCV2 = LOCV1 + N
LOCW = LOCV2 + N
C
C Check the workspace allocations.
CALL DCHKW( 'DSLUCS', LOCIW, LENIW, LOCW, LENW, IERR, ITER, ERR )
IF( IERR.NE.0 ) RETURN
C
IWORK(1) = LOCIL
IWORK(2) = LOCJL
IWORK(3) = LOCIU
IWORK(4) = LOCJU
IWORK(5) = LOCL
IWORK(6) = LOCDIN
IWORK(7) = LOCUU
IWORK(9) = LOCIW
IWORK(10) = LOCW
C
C Compute the Incomplete LU decomposition.
CALL DSILUS( N, NELT, IA, JA, A, ISYM, NL, IWORK(LOCIL),
$ IWORK(LOCJL), RWORK(LOCL), RWORK(LOCDIN), NU, IWORK(LOCIU),
$ IWORK(LOCJU), RWORK(LOCUU), IWORK(LOCNR), IWORK(LOCNC) )
C
C Perform the incomplete LU preconditioned
C BiConjugate Gradient Squared algorithm.
CALL DCGS(N, B, X, NELT, IA, JA, A, ISYM, DSMV,
$ DSLUI, ITOL, TOL, ITMAX, ITER, ERR, IERR, IUNIT,
$ RWORK(LOCR), RWORK(LOCR0), RWORK(LOCP),
$ RWORK(LOCQ), RWORK(LOCU), RWORK(LOCV1),
$ RWORK(LOCV2), RWORK, IWORK )
RETURN
C------------- LAST LINE OF DSLUCS FOLLOWS ----------------------------
END