Alison said:
>environment. In the Mac environment (unlike Unix and Microsoft),
>uninitialized memory is filled with garbage rather than zeros.
It may not be that it's a function of the Mac environment but of the Absoft
compiler. The Absoft compilers on other machines (and some other compilers)
may do the same thing.
See
Topic 5
for memo re: Installation of NONMEM with Powerstation Fortran.
Test Program for Powerstation and Microsoft Fortran
===================================
C Test program for Powerstation and Microsoft Fortran.
C From Alison Boeckmann, Nonmem Project Group, Jan. 1994
c Compile and run with each of three sets of options:
C /Ox
C /Ox /Op
C no optimization
C After each run, copy "testout.txt" to a uniquely-named file.
C All three output files should be identical.
PROGRAM TEST
INTEGER ICALL
INTEGER NEWIND
DOUBLE PRECISION THETA(5)
REAL DATREC(9)
INTEGER INDXS(20)
DOUBLE PRECISION F,G(10,10),H(10,10)
INTEGER NETAS,NEPS
DATA THETA/0.03,0.15,0.1,1.0,2/
DATA DATREC/102,1,1802,60,0.50,6.210,1,0,0/
NETAS=5
NEPS=1
ICALL=2
NEWIND=1
CALL PRED(ICALL,NEWIND,THETA,DATREC(1),INDXS,F,G,H)
OPEN (8,FILE='TESTOUT.TXT')
WRITE (8,999) F
WRITE (8,999) (G(I,1),I=1,NETAS),
X ((G(I,J+1),J=1,I),I=1,NETAS)
WRITE (8,999) (H(I,1),(H(I,J+1),J=1,NETAS),I=1,NEPS)
999 FORMAT (5(E16.8))
END
SUBROUTINE GETETA(ETA)
DIMENSION ETA(10)
DO 10 I=1,10
ETA(I)=0.0
10 CONTINUE
RETURN
END
SUBROUTINE SIMETA(ETA)
DOUBLE PRECISION ETA(10)
RETURN
END
SUBROUTINE SIMEPS(EPS)
DOUBLE PRECISION EPS(10)
RETURN
END
SUBROUTINE PRED (ICALL,NEWIND,THETA,DATREC,INDXS,F,G,H)
IMPLICIT DOUBLE PRECISION (A-Z)
REAL DATREC
SAVE
INTEGER ICALL,NEWIND,INDXS
DIMENSION THETA(*),DATREC(*),INDXS(*),G(10,*),H(10,*)
DIMENSION ETA(10),EPS(10)
COMMON/NMPRD1/IERPRD,NETEXT
COMMON/NMPRD2/ETEXT(3)
INTEGER IERPRD,NETEXT
CHARACTER*132 ETEXT
COMMON/ROCM12/MSEC
INTEGER MSEC
COMMON/NMPRD4/A1,AL1,A2,AL2,KA,T1,FF1,F1,FF2,F2,CT,Y,A00012
COMMON/NMPRD4/A00014,A00016,A00018,A00020,A00026,A00029,A00033
COMMON/NMPRD4/A00034,A00040,A00043,A00047,A00048,A00058,A00059
COMMON/NMPRD4/A00055,A00056,A00057,A00066,A00065,A00069,A00068
COMMON/NMPRD4/A00067,C00011,A00071,A00070,A00074,A00073,A00072
COMMON/NMPRD4/BBBBBB(0959)
IF (ICALL.EQ.4) THEN
IF (NEWIND.NE.2) CALL SIMETA(ETA)
CALL SIMEPS(EPS)
ELSE
IF (NEWIND.NE.2) THEN
CALL GETETA(ETA)
EPS(01)=0.D0
ENDIF
ENDIF
DOS=DATREC(04)
TIME=DATREC(05)
B00001=DEXP(ETA(01))
A1=THETA(01)*B00001
A00012=THETA(01)*B00001
B00002=DEXP(ETA(02))
AL1=THETA(02)*B00002
A00014=THETA(02)*B00002
B00003=DEXP(ETA(03))
A2=THETA(03)*B00003
A00016=THETA(03)*B00003
B00004=DEXP(ETA(04))
AL2=THETA(04)*B00004
A00018=THETA(04)*B00004
B00005=DEXP(ETA(05))
KA=THETA(05)*B00005
A00020=THETA(05)*B00005
IF(KA.LT.AL1)THEN
IERPRD=1
RETURN
ENDIF
IF(KA.LT.AL2)THEN
IERPRD=1
RETURN
ENDIF
T1=TIME
B00006=-AL1*T1
B00007=-KA*T1
B00008=DEXP(B00006)
B00009=DEXP(B00007)
FF1=B00008-B00009
A00021=-T1*A00014
A00022=-T1*A00020
A00023=B00008*A00021
A00024=B00009*A00022
A00026=-A00024
B00010=KA-AL1
F1=A1*FF1*KA/B00010
A00028=-A00014
B00011=FF1*KA/B00010
A00029=B00011*A00012
B00012=A1*KA/B00010
A00030=B00012*A00026
A00031=B00012*A00023
B00013=A1*FF1/B00010
A00032=B00013*A00020+A00030
B00014=-A1*FF1*KA/B00010/B00010
A00033=B00014*A00028+A00031
A00034=B00014*A00020+A00032
B00015=-AL2*T1
B00016=-KA*T1
B00017=DEXP(B00015)
B00018=DEXP(B00016)
FF2=B00017-B00018
A00035=-T1*A00018
A00036=-T1*A00020
A00037=B00017*A00035
A00038=B00018*A00036
A00040=-A00038
B00019=KA-AL2
F2=A2*FF2*KA/B00019
A00042=-A00018
B00020=FF2*KA/B00019
A00043=B00020*A00016
B00021=A2*KA/B00019
A00044=B00021*A00040
A00045=B00021*A00037
B00022=A2*FF2/B00019
A00046=B00022*A00020+A00044
B00023=-A2*FF2*KA/B00019/B00019
A00047=B00023*A00042+A00045
A00048=B00023*A00020+A00046
B00024=F1+F2
CT=DOS*B00024
A00052=A00048+A00034
A00055=DOS*A00043
A00056=DOS*A00047
A00057=DOS*A00052
A00058=DOS*A00029
A00059=DOS*A00033
Y=CT+CT*EPS(01)
A00060=A00059
A00061=A00058
A00062=A00057
A00063=A00056
A00064=A00055
A00065=EPS(01)*A00059+A00060
A00066=EPS(01)*A00058+A00061
A00067=EPS(01)*A00057+A00062
A00068=EPS(01)*A00056+A00063
A00069=EPS(01)*A00055+A00064
C00011=CT
A00070=A00059
A00071=A00058
A00072=A00057
A00073=A00056
A00074=A00055
G(01,1)=A00066
G(02,1)=A00065
G(03,1)=A00069
G(04,1)=A00068
G(05,1)=A00067
H(01,1)=C00011
H(01,02)=A00071
H(01,03)=A00070
H(01,04)=A00074
H(01,05)=A00073
H(01,06)=A00072
F=Y
RETURN
END
OOPS!
I said:
> David tried "Mac Fortran II v3.1.2 - Absoft - On a Macintosh IIci". The
> results of the program were incorrect with all compiler options that
> he tried, including no optimization at all.
David Bourne looked more closely and realized that the test program was
flawed. Certain variables are used which would have been initialized
to zero in the NONMEM environment, but were uninitialized in the test
environment. In the Mac environment (unlike Unix and Microsoft),
uninitialized memory is filled with garbage rather than zeros.
With correct initialization, Mac Fortran II results appear to be
correct. He is still checking various levels of optimization.
The bug in the driver does not explain the incorrect results from
Powerstation Fortran with /Ox, because these are in values that do
not depend on previous initialization.
Alison
Recently I sent this e-mail to nmusers mailing list:
> This is a message to people who use Microsoft Powerstation Fortran.
> Jaap Mandema brought to our attention a problem in a generated PRED
> routine when it is compiled with option /Ox (full optimization).
> The results are incorrect.
>
> I sent a little test program to Microsoft. The software developement
> group is working on it, but it is unlikely that they will have any
> bug fix or patch soon. The workaround is:
>
> Always use /Ox /Op options together.
> This should be done for all compiles: NONMEM, PREDPP, FSUBS.
> (NMTRAN does not need to be optimized.)
>
> I don't know if this bug affects the older Microsoft Fortran 5.x
> compilers.
>
> Would anyone like to try? If so, let me know and I'll send the little
> test program.
>
> -- Alison Boeckmann
Here is a follow-up.
Thanks to Bill Bachman, David Bourne, and Jaap Mandema, who gave this
a try.
The test program works correctly when compiled with Microsoft Fortran 5.1,
at all levels of optimization, including /Ox by itself. It is only the users
of Powerstation Fortran who should be careful not to use /Ox by itself.
David Bourne also tried the program on a Macintosh which has no
floating point chip. He tried "MS Fortran Compiler v5.1 IBM - with
Alternate Math Package Library - Running under SoftPC on a Macintosh
IIci". The results were correct with /Ox (and also with /Ox /Op, and
with no optimization.)
David tried "Mac Fortran II v3.1.2 - Absoft - On a Macintosh IIci". The
results of the program were incorrect with all compiler options that
he tried, including no optimization at all.
It appears that Mac Fortran II, with David's configuration, is flawed
and should not be used for NONMEM.
Alison