SUBROUTINE DJACG(MODE,M,N,Y,F, . FJAC,LDFJAC,YSCALE,FAC,IOPT, . WK,LWK,IWK,LIWK) c Copyright (c) 2006, Math a la Carte, Inc. c>> 2006-04-11 DJACG R. J. Hanson removed unneeded values saved. c>> 2003-07-08 DJACG R. J. Hanson fixed bug checking FAC(*) < 0. c>> 2003-07-08 DJACG R. J. Hanson fixed bug on MODE < 0 or > N. c>> 2002-06-21 DJACG R. J. Hanson modified Salane Code c--D replaces "?": ?JACG c Main routine DJACG, computing numerical derivatives. c ----------------------------------------------------------------- C C***BEGIN PROLOGUE DJACG C***DATE WRITTEN 860410 C***Date Modified 060411 C***CATEGORY NO. D4 C***KEYWORDS NUMERICAL DIFFERENCING, Numerical Jacobians, Numerical C Derivatives C***AUTHOR SALANE, DOUGLAS E., SANDIA NATIONAL LABORATORIES C NUMERICAL MATHEMATICS DIVISION,1642 C ALBUQUERQUE, NM 87185 C Interface reworked by Richard J. Hanson, Rice University. C Last changes made by Hanson after leaving Rice University. C C***PURPOSE SUBROUTINE DJACG C c Subroutine DJACG uses finite differences to compute the Jacobian for c a system of (M) equations in (N) variables. DJACG is designed c for use in numerical methods for solving nonlinear problems where a c Jacobian is evaluated repeatedly at neighboring arguments. c For example in a Gauss-Newton method for solving non-linear least squares c problems. DJACG is suitable for applications in which the Jacobian c is a dense rectangular matrix, over- exactly or under-determined. C The interface to the values of equations is with reverse c communication. The code DJACG saves and restores values between c calls. In particular DJACG is thread-safe. C The design allows for computation of derivatives that are a sum of C analytic and functional values that are not easily differentiated. C***DESCRIPTION C C SUBROUTINE PARAMETERS C C MODE.......An integer flag that directs user action with reverse C communication. Enter the routine with MODE=0. C DJACG returns with MODE=J when computing column J of C the Jacobian. When MODE=0 on return, the Jacobian C has been computed. C N..........[in] THE NUMBER OF VARIABLES. C M..........[in] THE NUMBER OF EQUATIONS. C Y(*).......[inout] AN ARRAY OF DIMENSION N. THE POINT AT WHICH THE C Jacobian IS TO BE EVALUATED. C F(*).......[in] AN ARRAY OF DIMENSION M. THE EQUATIONS EVALUATED AT C THE POINT Y. This must be defined when MODE .eq. 0 or C just after the first perturbed column is computed, C MODE .eq. 1. C FJAC(*,*)..[inout] AN ARRAY OF SIZE LDFJAC BY N. FJAC C CONTAINS THE Jacobian WHICH IS AN M BY N MATRIX. c Columns that are accumulated must have the analytic c part defined on entry or else be set to zero. c Columns that are skipped can be defined either before c or after the routine exits. C LDFJAC.....[in] THE LEADING DIMENSION OF THE ARRAY FJAC AS DIMENSIONED C IN THE CALLING ROUTINE. LDFJAC MUST BE GREATER OR C EQUAL TO M. C YSCALE(*)..[in] AN ARRAY OF LENGTH N. THE USER CAN PROVIDE C REPRESENTATIVE MAGNITUDES FOR Y VALUES IN THE ARRAY C YSCALE. THE USER CAN ALSO USE YSCALE TO PROVIDE C APPROPRIATE SIGNS FOR THE INCREMENTS. C FAC(*).....[inout] AN ARRAY OF LENGTH N. FAC(I) CONTAINS THE PERCENTAGE C FACTOR FOR DIFFERENCING. C c If FAC(1)=0., the array FAC is set to the square root of c machine unit roundoff on the first call to DJACG. c Unless the user wishes to initialize FAC, c the FAC array should not be altered between subsequent calls c to DJACG. c c The user may provide FAC values if desired. C If the user provides FAC values, FAC(i) should be c set to a value between 0 and 1. DJACG will not permit FAC(i) to c be set to a value that results in too small an increment. DJACG c ensures that c FACMIN <= FAC(i) <= FACMAX. C For further details on how the code chooses FACMIN and FACMAX see c the report [2]. c C C IOPT(*) [in] An integer array defining the methods used to C compute the derivatives. The default is to use one sided differences. C Entries in this array are interpreted as follows. C [0] Use the current settings for all remaining variables. The starting C setting is to simply use one sided differences. C [k > 0] Use the current settings for all variables from the last C specified up to and including variable $k$. C [-1] Set to use one sided differences. (Not needed at the C beginning since this is the default state.) C [-2] Set to use central differences. C [-3] Set to accumulate the result from whatever type of differences C have been specified above into initial values provided in FJAC. This C must be followed by a number $\geq 0$ as any number $< 0$ will turn C this off. C [-4] Skip variables. This must be followed by a value $\geq 0$ C as any number $< 0$ will turn this off. C Re. state -2: For variable p, compute numerical derivatives using C central divided differences. This will typically yield more C accuracy than the one-sided differences, but with the expense C of an additional function evaluation per variable. C The increment used for central differencing C is the default T=$macheps^{-1/6}$ C times the increment used in one-sided differencing. C To change this factor for succeeding C variables, assign a new value between calls with MODE .gt. 0 C in the storage location FAC(MODE). C The default value T=$macheps^{-1/6}$ is based on the C approximate relation T*FAC(J) =$macheps^{2/3}$. C This value is near an optimal choice, under certain conditions C on higher derivatives, provided FAC(J) = $macheps^{1/2}$. C Sometimes larger or smaller values of T will give more accuracy C than the default. C c For example if column 3 is not to be computed and column 4 is c to be accumulated, then IOPT(1)=2, followed by: C IOPT(2)=-4, IOPT(3)=3, IOPT(4)=-3, IOPT(5)=4, C IOPT(6)=-1, IOPT(7)=0 c c c wk(*)...[inout]a work array whose dimension is at least LWK. c LWK.....[in] the required length of the work array. LWK=3*M+18. c iwk(*)..[inout] an integer array whose dimension is at least LIWK. c The first 10 positions of IWK contain diagnostics c information. c c IWK(1) gives the number of times a function evaluation was computed. c c IWK(2) gives the number of columns in which three attempts were c made to increase a percentage factor for differencing (i.e. a c component in the FAC array) but the computed DEL remained c unacceptably small relative to Y(JCOL) or YSCALE(JCOL). In such c cases the percentage factor is set to the square root of the unit c roundoff of the machine. c c IWK(3) gives the number of columns in which the computed DEL was c zero to machine precision because Y(JCOL) or YSCALE(JCOL) was c zero. In such cases DEL is set to the square root of the unit c roundoff. c c IWK(4) gives the number of Jacobian columns which had to be c recomputed because the largest difference formed in the column c was close to zero relative to scale, where c c scale = max(|f (y)|,|f (y+del)|) c i i c c and i denotes the row index of the largest difference in the c column currently being processed. IWK(10) gives the last column c where this occurred. c c IWK(5) gives the number of columns whose largest difference is c close to zero relative to scale after the column has been c recomputed. c c IWK(6) gives the number of times scale information was not c available for use in the roundoff and truncation error tests. c this occurs when c c min (|f (y)|,|f (y+del)|) = 0. c i i c c where i is the index of the largest difference for the column c currently being processed. c c IWK(7) gives the number of times the increment for differencing c (DEL) was computed and had to be increased because (Y(JCOL)+DEL ) c -Y(JCOL)) was too small relative to Y(JCOL) or YSCALE(JCOL). c c IWK(8) gives the number of times a component of the FAC array was c reduced because changes in function values were large and excess c truncation error was suspected. IWK(9) gives the last column in c which this occurred. c c IWK(9) gives the index of the last column where the corresponding c component of the FAC array had to be reduced because excessive c truncation error was suspected. c c IWK(10) gives the index of the last column where the differeence c was small and the column had to be recomputed with an adjusted c increment (see iwk(4)). the largest derivative in this column c may be inaccurate due to excessive roundoff error. c c LIWK.[in] the length of the array iwk. LIWK >= 10+LNTBUF=21. c Using reverse communication, DJACG will return control to c the user each time a function value is required. The user must c provide the values of the function f evaluated at y where y is the c array return by DJACG. The user must set c c wk(i) = i-th component of f evaluated at y c c for i = 1,2,...,M. After wk is assigned, the user must recall DJACG. c When DJACG returns to the user with c c MODE = 0, c c the computation of the Jacobian is complete. c c Roundoff and truncation errors. c c Subroutine DJACG takes advantage of the way in which the Jacobian c is evaluated to adjust increments for differencing to control c roundoff and truncation errors. The routine usually (but not always) c requires one additional function evaluation to compute a column of the c Jacobian. Also, the routine returns a variety of error diagnostics c to warn users when computed derivatives may not be accurate. c c WARNING: DJACG does not guarantee the accuracy of the computed c derivatives. In order to save on function evaluations, heuristic c tecniques for increment adjustment and safeguarding increments are c used. These usually work well, but are not guaranteed. c c In order to determine the accuracy of computed derivates, users should c pay attention to the diagnostic array positions IWK(4), IWK(8), IWK(9) c and IWK(10). Non-zeros in these position indicate c possible large roundoff or truncation errors. Here is a summary. c C IWK(4) NONZERO => ROUNDOFF ERROR. IWK(10) LAST NOTED COLUMN. C IWK(8) NONZERO => TRUNCATION ERROR. IWK(9) LAST NOTED COLUMN. c c WARNING: Some of the diagnostics returned can only be interpreted c with a detailed knowledge of the routine. Nevertheless, they are c provided to give users full access to the information produced by c the subroutine. c c***references (1)D.E. Salane and L. F. Shampine c "An economical and efficient routine for computing c sparse Jacobians", Report no. SAND85-0977, c Sandia National Laboratories, Albuquerque,NM,87185. c (2)D.E. Salane c "Adaptive routines for forming Jacobians c numerically", Report no. SAND86-1319, Sandia c National Laboratories, Albuquerque, NM, 87185 c C***ROUTINES CALLED D1MACH C C REQUIRED FORTRAN INTRINSIC FUNCTIONS: C ABS,MAX,MIN,SIGN,SQRT C REQUIRED Math a la Carte FUNCTIONS: C D1MACH C C***END PROLOGUE DJACG C IMPLICIT NONE INTEGER MODE, N, M, LDFJAC, LWK, LIWK DOUBLE PRECISION Y(N),F(M),FJAC(LDFJAC,*),WK(LWK), . YSCALE(N),FAC(N) INTEGER IOPT(*),IWK(LIWK) C C SET CONSTANTS AND ALGORITHM PARAMETERS. C PERT........IS USED IN EXTRAPOLATION TEST. C FACMAX......FAC(I) <= FACMAX. C EXPFMN......FAC(I) => FACMIN WHERE FACMIN = U ** EXPFMN. C NIFAC.......ITERATION BOUND USED IN COMPUTING INCREMENT. EXTERNAL D1MACH DOUBLE PRECISION D1MACH DOUBLE PRECISION PERT, FACMAX DOUBLE PRECISION USQT, UEGT, UMEGT, UQRT, U3QRT, . U7EGT, FACMIN, SAVY, AY, SGN, CFACT, . DEL, DELM, T1, T2, SAVDEL, DMAX, RDEL, ADIFF, . DIFF, SF, SDF, FJACL, FMJ, DFMJ, RMXFDF, RMNFDF, U INTEGER NIFAC INTEGER I, J, ITRY, IRDEL, IRCMP, IROWMX, IROW INTEGER JCOL, NS, STATEA, STATED, STATES INTEGER LRBUF, LNTBUF PARAMETER(LRBUF=18, LNTBUF=11) DOUBLE PRECISION RBUF(LRBUF) INTEGER INTBUF(LNTBUF) C C***FIRST EXECUTABLE STATEMENT DJACG IF(MODE .gt. 0 .and. MODE .le. N) GO TO 90 C Do error checking on some parameters. A returned value of MODE < 0 C denotes that argument number -MODE has an error condition associated C with it. C SUBROUTINE DJACG(MODE,M,N,Y,F, C . FJAC,LDFJAC,YSCALE,FAC,IOPT, C . WK,LWK,IWK,LIWK) C Variables 1-5, 6-10, 11-14 C Must have MODE=0 the first time. IF(MODE .lt. 0 .or. MODE .gt. N) THEN C This says that MODE < 0 or MODE > N is error. MODE = 1 GO TO 1000 END IF MODE=2 C Must have M > 0 IF(M .le. 0) GO TO 1000 MODE=3 C Must have N > 0 IF(N .le. 0) GO TO 1000 MODE=7 C Must have LDFJAC .ge. M if N > 1. IF(LDFJAC .lt. M .and. N .gt. 1) GO TO 1000 C Must have FAC() >= 0. The value FAC(1)=0. C skips this test on the rest of FAC(*). IF(FAC(1) .ne. 0.) THEN MODE=9 DO 1010 J=1,N IF(FAC(J) .lt. 0.) GO TO 1000 1010 CONTINUE END IF MODE=10 C Run throught the IOPT() array making sure that the C state changes and indices make sense. NS=1 1070 CONTINUE C A positive value .gt. N is an error. IF(IOPT(NS).gt. N) GO TO 1000 C A value of 0 (or N) says to use the present setting for C the rest of the variables IF(IOPT(NS).eq. 0 .or. IOPT(NS) .eq. N) GO TO 1080 C A postitive value < N must be followed by a zero, N, or a C state change. IF(IOPT(NS).gt. 0) THEN NS=NS+1 GO TO 1070 END IF IF(IOPT(NS) .lt. -4) GO TO 1000 IF(IOPT(NS) .ge. -2) THEN NS=NS+1 GO TO 1070 END IF C An accumulation or skip must be followed by a non-negative index. IF(IOPT(NS+1) .lt. 0) GO TO 1000 NS=NS+1 GO TO 1070 1080 CONTINUE MODE=12 C Must have LWK >= 3*M+LRBUF IF(LWK .lt. 3*M+LRBUF) GO TO 1000 MODE=14 C Must have LIWK >= 21 IF(LIWK .lt. 21) GO TO 1000 C c Since reverse communication, initialize JCOL only on first call. JCOL=0 C C COMPUTE APPROPRIATE MACHINE CONSTANTS. C U = D1MACH(4) USQT = SQRT(U) c UQRT = U**P25 c U3QRT = U**P75 c UEGT = U**P125 c U7EGT = U**P875 c FACMIN = U**EXPFMN UQRT = SQRT(USQT) T1=1.D0/USQT U3QRT = U/UQRT UEGT = SQRT(UQRT) UMEGT = 1.D0/UEGT U7EGT = U*UMEGT FACMIN = U3QRT PERT=2.0D+0 FACMAX=1.D-1 NIFAC=3 DO 10 J = 1,LIWK IWK(J) = 0 10 CONTINUE C The value FAC=0. means that this code wants to initialize. IF(FAC(1) .eq. 0.D0) THEN DO 20 J = 1,N FAC(J) = USQT 20 CONTINUE END IF C Default to not accumulate: STATEA=1 C Default to use one-sided differences. STATED=1 CFACT=0.D0 C Default to not skip computing columns. STATES=1 C Index for processing IOPT(*) states and changes. NS = 1 C 30 CONTINUE JCOL = JCOL + 1 C IF IOPT(NS) .eq. 0, the current states hold for the rest C of the columns. 1020 CONTINUE IF(IOPT(NS) .ne. 0) THEN C A state change will occur: IF(JCOL .eq. IOPT(NS)+1) THEN C This signals a state change. Namely the last variable C in the group has been processed and a change is required. NS=NS+1 END IF IF(IOPT(NS) .lt. 0) THEN C Set for one-sided or central differences. IF(IOPT(NS) .ge. -2) THEN STATED=-IOPT(NS) STATEA=1 ELSE C Set for accumulation or skipping variables. IF(IOPT(NS) .eq. -3) STATEA=2 IF(IOPT(NS) .eq. -4) THEN STATES=2 STATEA=1 END IF END IF NS=NS+1 ELSE GO TO 1030 END IF C Still in range for a state, without change to that state. GO TO 1020 END IF 1030 CONTINUE IF(STATED .gt. 1) THEN IF(CFACT .eq. 0D0) THEN C CFACT=U**(-1.D0/6.D0) C Compute U**(-1/6) using Newton's method; compute the cube C root of U**(-1/2). This is to avoid the ** elmentary function C which may not be thread-safe or re-entrant. CFACT = UMEGT*UMEGT DO 1050 J=1,14 CFACT=(2.D0*CFACT+T1/(CFACT*CFACT))/3.D0 C Converged, to over 7/8 precision. Escapes from loop. IF(ABS(CFACT*CFACT*CFACT*USQT - 1.D0) .lt. U7EGT) . GO TO 1060 1050 CONTINUE 1060 CONTINUE END IF END IF C Check if this column is to be skipped. IF(STATES .gt. 1) GO TO 170 IRCMP = 0 ITRY = 0 C Check if this column is to be accumulated. IF(STATEA .gt. 1) THEN C Save the initial contents of the array column for final accumulation. DO 40 IROW=1,M WK(IROW+2*M)=FJAC(IROW,JCOL) 40 CONTINUE END IF C 50 CONTINUE SAVY = Y(JCOL) C c Compute DEL. If DEL is too small increase FAC and recompute c DEL. NIFAC attempts are made to increase FAC and find an c appropriate DEL. If DEL can't be found in this manner, DEL is computed c with FAC set to the square root of the machine precision(=USQT). c If DEL is zero to machine precision because T is zero or c YSCALE is zero, DEL is set to USQT. C IRDEL = 0 IF (YSCALE(1) .ne. 0.D0) THEN AY = ABS(YSCALE(JCOL)) ELSE AY = ABS(Y(JCOL)) END IF SGN=1.D0 IF(YSCALE(1) .ne. 0.D0)SGN = SIGN(1.D0,YSCALE(JCOL)) DELM = U7EGT*AY C DO 60 J = 1,NIFAC DEL = FAC(JCOL)*AY*SGN IF (DEL.EQ.0.D0) THEN DEL = USQT*SGN IF (ITRY.EQ.0) IWK(3) = IWK(3) + 1 END IF T1 = Y(JCOL) + DEL DEL = T1 - Y(JCOL) IF (ABS(DEL).LT.DELM) THEN IF (J.GE.NIFAC) GO TO 60 IF (IRDEL.EQ.0) THEN IRDEL = 1 IWK(7) = IWK(7) + 1 END IF T1 = FAC(JCOL)*UMEGT FAC(JCOL) = MIN(T1,FACMAX) ELSE GO TO 70 END IF 60 CONTINUE C FAC(JCOL) = USQT DEL = USQT*AY*SGN IWK(2) = IWK(2) + 1 C 70 CONTINUE C See if central divided differences are intended. IF(STATED .gt. 1) THEN DEL = CFACT*DEL END IF SAVDEL = DEL Y(JCOL) = Y(JCOL) + DEL IWK(1) = IWK(1) + 1 C c With reverse communication, DJACG returns to calling routine c for function values. Afterwards, DJACG transfers control c to the statement after the comment - ENTRY POINT FOR REVERSE C COMMUNICATION. C Pack up scalars that need saving. Place in store within C WK(*) and IWK(*). GO TO 190 80 CONTINUE MODE = JCOL RETURN C ENTRY POINT FOR REVERSE COMMUNICATON. 90 CONTINUE C Unpack scalars that needed saving. C Take from WK(*) and IWK(*) where they are stored. GO TO 220 100 CONTINUE C See if this is the second evaluation for a central difference. C This is true STATED .gt. 1 .and. ITRY .eq. 1. C and ITRY = 1. RDEL = 1.D0/SAVDEL IF(STATED .eq. 1) GO TO 140 C IF ITRY = 1, proceed and compute the central divided difference. IF(ITRY .eq. 1) GO TO 120 C Otherwise ITRY = 0, so an alternate evaluation is needed. Y(JCOL)=SAVY-DEL DO 110 I=1,M WK(I+M)=WK(I) 110 CONTINUE ITRY=1 IWK(1) = IWK(1) + 1 GO TO 190 120 CONTINUE C Compute the central divided difference. RDEL=.5*RDEL DO 130 I=1,M FJAC(I,JCOL)=RDEL*(WK(I+M)-WK(I)) 130 CONTINUE C Restore the component of Y() and move on. Y(JCOL) = SAVY ITRY=0 GO TO 170 C Restore the component of Y() and move on. 140 CONTINUE Y(JCOL) = SAVY C c Compute the Jacobian entries. c Use largest elements in a column to determine scaling c information for roundoff and truncation error tests. C DMAX = 0.D0 IROWMX = 1 C DO 150 IROW = 1,M DIFF = WK(IROW) - F(IROW) ADIFF = ABS(DIFF) IF (ADIFF.GE.DMAX) THEN IROWMX = IROW DMAX = ADIFF SF = F(IROW) SDF = WK(IROW) END IF FJACL = DIFF*RDEL IF (ITRY.EQ.1) THEN WK(IROW) = FJACL WK(IROW+M) = FJAC(IROW,JCOL) END IF FJAC(IROW,JCOL) = FJACL 150 CONTINUE c If a column is being recomputed (ITRY=1), this section of the c code performs an extrapolation test to enable the code to c compute small derivatives more accurately. C IF (ITRY.EQ.1) THEN T1 = WK(IROWMX+M) T2 = WK(IROWMX)*FAC(JCOL) IF (ABS(T2).LT.ABS(T1)*PERT) THEN T1 = FAC(JCOL)*FAC(JCOL) FAC(JCOL) = MAX(T1,FACMIN) DO 160 IROW = 1,M FJACL = WK(IROW+M) FJAC(IROW,JCOL) = FJACL 160 CONTINUE END IF END IF C FMJ = ABS(SF) DFMJ = ABS(SDF) RMXFDF = MAX(FMJ,DFMJ) RMNFDF = MIN(FMJ,DFMJ) C c If scale information is not available, perform no roundoff c or truncation error tests. C IF (RMNFDF.NE.0.D0) THEN C c Test for possible roundoff error (first test) c and also for possible serious roundoff error (second test). C IF (DMAX.LE. (U3QRT*RMXFDF)) THEN IF (DMAX.LE. (U7EGT*RMXFDF)) THEN IF (ITRY.EQ.0) THEN T1 = SQRT(FAC(JCOL)) FAC(JCOL) = MIN(T1,FACMAX) IRCMP = 1 IWK(4) = IWK(4) + 1 IWK(10) = JCOL ELSE IWK(5) = IWK(5) + 1 END IF ELSE T1 = UMEGT*FAC(JCOL) FAC(JCOL) = MIN(T1,FACMAX) END IF END IF C c Test for possible truncation error. C IF (DMAX.GT.UQRT*RMXFDF) THEN T1 = FAC(JCOL)*UEGT FAC(JCOL) = MAX(T1,FACMIN) IWK(8) = IWK(8) + 1 IWK(9) = JCOL END IF ELSE IWK(6) = IWK(6) + 1 C END IF C c If serious roundoff error is suspected, recompute the c column. C C IF (IRCMP.EQ.1) THEN IRCMP = 0 ITRY = 1 GO TO 50 END IF ITRY = 0 C Branch to this place if this column is skipped. 170 CONTINUE C May have asked for this column to be accumulated. IF(STATEA .gt. 1 .and. STATES .eq. 1) THEN DO 180 IROW=1,M FJAC(IROW,JCOL)=FJAC(IROW,JCOL)+WK(IROW+2*M) 180 CONTINUE END IF IF (JCOL.LT.N) GO TO 30 C This is the signal that all Jacobian columns are computed. MODE = 0 RETURN C Internal procedure to save local data. C This step is needed to make the code re-entrant and threadsafe. 190 CONTINUE RBUF(01)=FACMIN RBUF(02)=FACMAX RBUF(03)=CFACT RBUF(04)=USQT RBUF(05)=UEGT RBUF(06)=UMEGT RBUF(07)=UQRT RBUF(08)=U3QRT RBUF(09)=U7EGT RBUF(10)=U RBUF(11)=CFACT RBUF(12)=SAVDEL RBUF(13)=PERT RBUF(14)=SAVY RBUF(15)=AY RBUF(16)=SGN RBUF(17)=DEL RBUF(18)=DELM C Save data values in calling program unit. When the code is C entered again these are extracted from storage the store C in the calling program. This eliminates the need to save C any variables. DO 200 I=1,LRBUF WK(I+3*M)=RBUF(I) 200 CONTINUE INTBUF(01)=NIFAC INTBUF(02)=I INTBUF(03)=J INTBUF(04)=ITRY INTBUF(05)=IRDEL INTBUF(06)=IRCMP INTBUF(07)=JCOL INTBUF(08)=NS INTBUF(09)=STATEA INTBUF(10)=STATED INTBUF(11)=STATES C Save data values in calling program unit. When the code is C called again these are extracted. DO 210 I =1,LNTBUF IWK(I+10)=INTBUF(I) 210 CONTINUE GO TO 80 C Internal procedure to unpack saved data. 220 CONTINUE C Extract saved data from store in the calling program unit. DO 230 I=1,LRBUF RBUF(I)= WK(I+3*M) 230 CONTINUE FACMIN=RBUF(01) FACMAX=RBUF(02) CFACT =RBUF(03) USQT =RBUF(04) UEGT =RBUF(05) UMEGT =RBUF(06) UQRT =RBUF(07) U3QRT =RBUF(08) U7EGT =RBUF(09) U =RBUF(10) CFACT =RBUF(11) SAVDEL=RBUF(12) PERT =RBUF(13) SAVY =RBUF(14) AY =RBUF(15) SGN =RBUF(16) DEL =RBUF(17) DELM =RBUF(18) C Extract saved data from store in the calling program unit. DO 240 I =1,LNTBUF INTBUF(I)=IWK(I+10) 240 CONTINUE NIFAC =INTBUF(01) I =INTBUF(02) J =INTBUF(03) ITRY =INTBUF(04) IRDEL =INTBUF(05) IRCMP =INTBUF(06) JCOL =INTBUF(07) NS =INTBUF(08) STATEA=INTBUF(09) STATED=INTBUF(10) STATES=INTBUF(11) GO TO 100 1000 CONTINUE C Error return when arguments are invalid. MODE=-MODE END .