C********************************************************************** C C Copyright (C) 1992 Roland W. Freund and Noel M. Nachtigal C All rights reserved. C C This code is part of a copyrighted package. For details, see the C file "cpyrit.doc" in the top-level directory. C C ***************************************************************** C ANY USE OF THIS CODE CONSTITUTES ACCEPTANCE OF THE TERMS OF THE C COPYRIGHT NOTICE C ***************************************************************** C C********************************************************************** C C This file contains the routine for the QMR algorithm for C symmetric matrices, using the three-term recurrence variant of C the Lanczos algorithm without look-ahead. C C********************************************************************** C SUBROUTINE DSQMX (NDIM,NLEN,NLIM,VECS,TOL,INFO) C C Purpose: C This subroutine uses the QMR algorithm to solve linear systems. C It runs the algorithm to convergence or until a user-specified C limit on the number of iterations is reached. It is set up to C solve symmetric systems starting with identical starting vectors. C C The code is set up to solve the system A x = b with initial C guess x_0 = 0. Here A x = b denotes the preconditioned system, C and it is connected with the original system as follows. Let C B y = c be the original unpreconditioned system to be solved, and C let y_0 be an arbitrary initial guess for its solution. Then: C A x = b, where A = M_1^{-1} B M_2^{-1}, C x = M_2 (y - y_0), b = M_1^{-1} (c - B y_0). C Here M = M_1 M_2 is the preconditioner. C C To recover the final iterate y_n for the original system B y = c C from the final iterate x_n for the preconditioned system A x = b, C set C y_n = y_0 + M_2^{-1} x_n. C C The implementation does not have look-ahead, so it is less robust C than the full version. C C Parameters: C For a description of the parameters, see the file "dsqmx.doc" in C the current directory. C C External routines used: C double precision dlamch(ch) C LAPACK routine, computes machine-related constants. C double precision dnrm2(n,x,incx) C BLAS-1 routine, computes the 2-norm of x. C subroutine daxpby(n,z,a,x,b,y) C Library routine, computes z = a * x + b * y. C double precision ddot(n,x,incx,y,incy) C BLAS-1 routine, computes y^H * x. C double precision zqmxom(n) C User-supplied routine, specifies the QMR scaling factors. C subroutine drandn(n,x,seed) C Library routine, fills x with random numbers. C subroutine drotg(a,b,cos,sin) C BLAS-1 routine, computes the Givens rotation which rotates the C vector [a; b] into [ sqrt(a**2 + b**2); 0 ]. C C Noel M. Nachtigal C May 25, 1993 C C********************************************************************** C INTRINSIC DABS, DBLE, DMAX1, DSQRT, MAX0 INTRINSIC MOD EXTERNAL DLAMCH, DNRM2, DAXPBY, DDOT, DRANDN, DROTG, DSQMXO DOUBLE PRECISION DDOT DOUBLE PRECISION DLAMCH, DNRM2, DSQMXO C INTEGER INFO(4), NDIM, NLEN, NLIM DOUBLE PRECISION VECS(NDIM,7) DOUBLE PRECISION TOL C C Miscellaneous parameters. C DOUBLE PRECISION DHUN, DONE, DTEN, DZERO PARAMETER (DHUN = 1.0D2,DONE = 1.0D0,DTEN = 1.0D1,DZERO = 0.0D0) C C Local variables, permanent. C INTEGER IERR, ISN, ISNM1, ISNM2, IVN, IVNM1, IVNP1, IWN, IWNM1 SAVE IERR, ISN, ISNM1, ISNM2, IVN, IVNM1, IVNP1, IWN, IWNM1 INTEGER IWNP1, N, RETLBL, TF, TRES, VF SAVE IWNP1, N, RETLBL, TF, TRES, VF DOUBLE PRECISION DN, DNM1, DNP1, RHSN, SCSN, SCSNM1, SINN, SINNM1 SAVE DN, DNM1, DNP1, RHSN, SCSN, SCSNM1, SINN, SINNM1 DOUBLE PRECISION COSN, COSNM1, CSIN, CSINP1, MAXOMG, OMGN SAVE COSN, COSNM1, CSIN, CSINP1, MAXOMG, OMGN DOUBLE PRECISION OMGNP1, R0, RESN, RHON, RHONP1, SCVN, SCVNP1 SAVE OMGNP1, R0, RESN, RHON, RHONP1, SCVN, SCVNP1 DOUBLE PRECISION SCWN, SCWNP1, TMAX, TMIN, TNRM, UCHK, UNRM SAVE SCWN, SCWNP1, TMAX, TMIN, TNRM, UCHK, UNRM C C Local variables, transient. C INTEGER INIT, REVCOM DOUBLE PRECISION RHN, RHNM1, RHNM2, RHNP1, RHSNP1 DOUBLE PRECISION DTMP, SCVNM1, SCWNM1 C C Initialize some of the permanent variables. C DATA RETLBL /0/ C C Check the reverse communication flag to see where to branch. C REVCOM RETLBL Comment C 0 0 first call, go to label 10 C 1 30 returning from AXB, go to label 30 C 1 60 returning from AXB, go to label 60 C 2 40 returning from ATXB, go to label 40 C REVCOM = INFO(2) INFO(2) = 0 IF (REVCOM.EQ.0) THEN N = 0 IF (RETLBL.EQ.0) GO TO 10 ELSE IF (REVCOM.EQ.1) THEN IF (RETLBL.EQ.30) THEN GO TO 30 ELSE IF (RETLBL.EQ.60) THEN GO TO 60 END IF ELSE IF (REVCOM.EQ.2) THEN IF (RETLBL.EQ.40) GO TO 40 END IF IERR = 1 GO TO 80 C C Check whether the inputs are valid. C 10 IERR = 0 IF (NDIM.LT.1) IERR = 2 IF (NLEN.LT.1) IERR = 2 IF (NLIM.LT.1) IERR = 2 IF (NLEN.GT.NDIM) IERR = 2 IF (IERR.NE.0) GO TO 80 C C Extract from INFO the output units TF and VF, the true residual C flag TRES, and the left starting vector flag INIT. C VF = MAX0(INFO(1),0) INIT = VF / 100000 VF = VF - INIT * 100000 TRES = VF / 10000 VF = VF - TRES * 10000 TF = VF / 100 VF = VF - TF * 100 C C Extract and check the various tolerances. C TNRM = DLAMCH('E') * DTEN TMIN = DSQRT(DSQRT(DLAMCH('S'))) TMAX = DONE / TMIN IF (TOL.LE.DZERO) TOL = DSQRT(DLAMCH('E')) C C Start the trace messages and convergence history. C IF (VF.NE.0) WRITE (VF,'(I8,2E11.4)') 0, DONE, DONE IF (TF.NE.0) WRITE (TF,'(I8,2E11.4)') 0, DONE, DONE C C Initialize the wrapped indices. C ISNM1 = 5 ISN = ISNM1 IVN = 3 IVNP1 = IVN IWN = 3 IWNP1 = IWN C C Set x_0 = 0 and compute the norm of the initial residual. C CALL DAXPBY (NLEN,VECS(1,IVN),DONE,VECS(1,2),DZERO,VECS(1,IVN)) CALL DAXPBY (NLEN,VECS(1,1),DZERO,VECS(1,1),DZERO,VECS(1,1)) R0 = DNRM2(NLEN,VECS(1,IVN),1) IF ((TOL.GE.DONE).OR.(R0.EQ.DZERO)) GO TO 80 C C Check whether the auxiliary vector must be supplied. C CSYM IF (INIT.EQ.0) CALL DRANDN (NLEN,VECS(1,IWN),1) C C Compute scale factors and check for invariant subspaces. C SCVNP1 = R0 CSYM SCWNP1 = DNRM2(NLEN,VECS(1,IWN),1) SCWNP1 = SCVNP1 IF (SCVNP1.LT.TNRM) IERR = IERR + 16 IF (IERR.NE.0) GO TO 80 DNP1 = DDOT(NLEN,VECS(1,IWN),1,VECS(1,IVN),1) / ( SCVNP1 * SCWNP1 \$) IF ((SCVNP1.GE.TMAX).OR.(SCVNP1.LE.TMIN)) THEN DTMP = DONE / SCVNP1 CALL DAXPBY (NLEN,VECS(1,IVN),DTMP,VECS(1,IVN),DZERO,VECS(1,IVN \$)) SCVNP1 = DONE END IF IF ((SCWNP1.GE.TMAX).OR.(SCWNP1.LE.TMIN)) THEN DTMP = DONE / SCWNP1 CSYM CALL DAXPBY (NLEN,VECS(1,IWN),DTMP,VECS(1,IWN),DZERO,VECS(1,IWN)) SCWNP1 = DONE END IF RHONP1 = SCVNP1 CSINP1 = SCWNP1 SCVNP1 = DONE / SCVNP1 SCWNP1 = DONE / SCWNP1 C C Initialize the variables. C N = 1 DN = DONE COSN = DONE RESN = DONE COSNM1 = DONE OMGN = DZERO SCSN = DZERO SCVN = DZERO SCWN = DZERO SCSNM1 = DZERO SINN = DZERO SINNM1 = DZERO OMGNP1 = DSQMXO(N) RHSN = OMGNP1 * R0 MAXOMG = DONE / OMGNP1 C C This is one step of the classical Lanczos algorithm. C 20 IVNM1 = IVN IVN = IVNP1 IVNP1 = MOD(N,2) + 3 IWNM1 = IWN IWN = IWNP1 IWNP1 = MOD(N,2) + 3 C C Check whether D_n is nonsingular. C DNM1 = DN DN = DNP1 IF (DN.EQ.DZERO) THEN IERR = 8 GO TO 80 END IF C C Have the caller carry out AXB, then return here. C CALL AXB (VECS(1,IVN),VECS(1,7)) C INFO(2) = 1 INFO(3) = IVN INFO(4) = 7 RETLBL = 30 RETURN 30 RETLBL = 0 C C Compute H_{n-1,n} and build part of the vector v_{n+1}. C SCVNM1 = SCVN CSIN = CSINP1 SCVN = SCVNP1 DTMP = CSIN * DN / DNM1 * SCVNM1 / SCVN CALL DAXPBY (NLEN,VECS(1,IVNP1),DONE,VECS(1,7),-DTMP,VECS(1,IVNM1) \$) C C Have the caller carry out ATXB, then return here. C CALL ATXB (VECS(1,IWN),VECS(1,7)) C INFO(2) = 2 INFO(3) = IWN INFO(4) = 7 RETLBL = 40 CSYM RETURN 40 RETLBL = 0 C C Build part of the vector w_{n+1}. C SCWNM1 = SCWN RHON = RHONP1 SCWN = SCWNP1 DTMP = RHON * DN / DNM1 * SCWNM1 / SCWN CSYM CALL DAXPBY (NLEN,VECS(1,IWNP1),DONE,VECS(1,7),-DTMP,VECS(1,IWNM1)) C C Compute H_{nn} and finish the new vectors. C RHN = SCVN * SCWN * DDOT(NLEN,VECS(1,IWN),1,VECS(1,IVNP1),1) / DN CALL DAXPBY (NLEN,VECS(1,IVNP1),DONE,VECS(1,IVNP1),-RHN,VECS(1,IVN \$)) CSYM CALL DAXPBY (NLEN,VECS(1,IWNP1),DONE,VECS(1,IWNP1),-RHN,VECS(1,IWN)) C C Compute scale factors and check for invariant subspaces. C IERR = 0 SCVNP1 = DNRM2(NLEN,VECS(1,IVNP1),1) CSYM SCWNP1 = DNRM2(NLEN,VECS(1,IWNP1),1) SCWNP1 = SCVNP1 RHONP1 = SCVN * SCVNP1 CSINP1 = SCWN * SCWNP1 RHNP1 = RHONP1 IF (SCVNP1.LT.TNRM) IERR = IERR + 16 IF (IERR.NE.0) GO TO 50 DNP1 = DDOT(NLEN,VECS(1,IWNP1),1,VECS(1,IVNP1),1) / ( SCVNP1 * \$SCWNP1 ) IF ((SCVNP1.GE.TMAX).OR.(SCVNP1.LE.TMIN)) THEN DTMP = DONE / SCVNP1 CALL DAXPBY (NLEN,VECS(1,IVNP1),DTMP,VECS(1,IVNP1),DZERO,VECS(1 \$,IVNP1)) SCVNP1 = DONE END IF IF ((SCWNP1.GE.TMAX).OR.(SCWNP1.LE.TMIN)) THEN DTMP = DONE / SCWNP1 CSYM CALL DAXPBY (NLEN,VECS(1,IWNP1),DTMP,VECS(1,IWNP1),DZERO,VECS(1,IWNP1)) SCWNP1 = DONE END IF SCVNP1 = DONE / SCVNP1 SCWNP1 = DONE / SCWNP1 C C The QMR code starts here. C Multiply the new column by the previous omegas. C Get the next scaling factor omega(i) and update MAXOMG. C 50 RHNM1 = CSIN * DN * OMGN / DNM1 OMGN = OMGNP1 RHN = OMGN * RHN OMGNP1 = DSQMXO(N+1) RHNP1 = OMGNP1 * RHNP1 MAXOMG = DMAX1(MAXOMG,DONE/OMGN) C C Apply the previous rotations. C RHNM2 = SINNM1 * RHNM1 RHNM1 = COSNM1 * RHNM1 COSNM1 = COSN SINNM1 = SINN DTMP = RHNM1 RHNM1 = COSNM1 * DTMP + SINNM1 * RHN RHN = -SINNM1 * DTMP + COSNM1 * RHN C C Compute the rotation for the last element (this also applies it). C CALL DROTG (RHN,RHNP1,COSN,SINN) C C Apply the new rotation to the right-hand side vector. C RHSNP1 = -SINN * RHSN RHSN = COSN * RHSN C C Compute the next search direction s_n. C ISNM2 = ISNM1 ISNM1 = ISN ISN = MOD(N-1,2) + 5 DTMP = SCSNM1 * RHNM2 / SCVN CALL DAXPBY (NLEN,VECS(1,ISN),DONE,VECS(1,IVN),-DTMP,VECS(1,ISNM2) \$) SCSNM1 = SCSN DTMP = SCSNM1 * RHNM1 / SCVN CALL DAXPBY (NLEN,VECS(1,ISN),DONE,VECS(1,ISN),-DTMP,VECS(1,ISNM1) \$) SCSN = SCVN / RHN C C Compute the new QMR iterate, then scale the search direction. C DTMP = SCSN * RHSN CALL DAXPBY (NLEN,VECS(1,1),DONE,VECS(1,1),DTMP,VECS(1,ISN)) DTMP = DABS(SCSN) IF ((DTMP.GE.TMAX).OR.(DTMP.LE.TMIN)) THEN CALL DAXPBY (NLEN,VECS(1,ISN),SCSN,VECS(1,ISN),DZERO,VECS(1,ISN \$)) SCSN = DONE END IF C C Compute the residual norm upper bound. C If the scaled upper bound is within one order of magnitude of the C target convergence norm, compute the true residual norm. C RHSN = RHSNP1 UNRM = DSQRT(DBLE(N+1)) * MAXOMG * DABS(RHSNP1) / R0 UCHK = UNRM IF ((TRES.EQ.0).AND.(UNRM/TOL.GT.DTEN).AND.(N.LT.NLIM)) GO TO 70 C C Have the caller carry out AXB, then return here. C CALL AXB (VECS(1,1),VECS(1,7)) C INFO(2) = 1 INFO(3) = 1 INFO(4) = 7 RETLBL = 60 RETURN 60 RETLBL = 0 CALL DAXPBY (NLEN,VECS(1,7),DONE,VECS(1,2),-DONE,VECS(1,7)) RESN = DNRM2(NLEN,VECS(1,7),1) / R0 UCHK = RESN C C Output the trace messages and convergence history. C 70 IF (VF.NE.0) WRITE (VF,'(I8,2E11.4)') N, UNRM, RESN IF (TF.NE.0) WRITE (TF,'(I8,2E11.4)') N, UNRM, RESN C C Check for convergence or termination. Stop if: C 1. algorithm converged; C 2. there is an error condition; C 3. the residual norm upper bound is smaller than the computed C residual norm by a factor of at least 100; C 4. algorithm exceeded the iterations limit. C IF (RESN.LE.TOL) THEN IERR = 0 GO TO 80 ELSE IF (IERR.NE.0) THEN GO TO 80 ELSE IF (UNRM.LT.UCHK/DHUN) THEN IERR = 4 GO TO 80 ELSE IF (N.GE.NLIM) THEN IERR = 4 GO TO 80 END IF C C Update the running counter. C N = N + 1 GO TO 20 C C Done. C 80 NLIM = N RETLBL = 0 INFO(1) = IERR C RETURN END C C********************************************************************** C DOUBLE PRECISION FUNCTION DSQMXO (I) C C Purpose: C Returns the scaling parameter OMEGA(I). C C Parameters: C I = the index of the parameter OMEGA (input). C C Noel M. Nachtigal C March 30, 1993 C C********************************************************************** C INTEGER I C DSQMXO = 1.0D0 C RETURN END C C**********************************************************************