forums.silverfrost.com

[Ananda]

I have identified a problem with the OPTIMISE switch with FTN95. The following screen dump demonstrate this.

------------------------------------------------------------
D:contramc7tubaver16opt_test>type test.inc
integer*4 M_vtype,M_Ftype,N_vtype,N_Ftype,N_Evlyrs
parameter (M_vtype=10)
parameter (M_Ftype=3)
real*4 dmfc_allyr(M_vtype,M_Ftype),dsfc_allyr(M_vtype,M_Ftype)
common /cdata/N_vtype,N_Ftype,N_Evlyrs,dmfc_allyr,dsfc_allyr

D:contramc7tubaver16opt_test>type test.for
C*******************************************************************************
PROGRAM TEST
C*******************************************************************************
include 'test.inc'

c integer*4 M_vtype,M_Ftype,N_vtype,N_Ftype,N_Evlyrs
c parameter (M_vtype=10)
c parameter (M_Ftype=3)
c real*4 dmfc_allyr(M_vtype,M_Ftype),dsfc_allyr(M_vtype,M_Ftype)
c common /cdata/N_vtype,N_Ftype,N_Evlyrs,dmfc_allyr,dsfc_allyr

integer*4 S_dmfc,A_dmfc,A_dsfc
N_vtype=1
N_Ftype=2
N_Evlyrs=1
S_dmfc=8*N_vtype*N_Ftype*N_Evlyrs
call get_storage@(A_dmfc,s_dmfc)
call get_storage@(A_dsfc,s_dmfc)
call fuel_con_out(DCORE8(A_dmfc),DCORE8(A_dsfc))
call return_storage@(A_dsfc)
call return_storage@(A_dmfc)
end
C*******************************************************************************
subroutine fuel_con_out(dm_fuel_con,ds_fuel_con)
C*******************************************************************************

include 'test.inc'
c
real*8 dm_fuel_con(N_vtype,N_Evlyrs,N_Ftype)
real*8 ds_fuel_con(N_vtype,N_Evlyrs,N_Ftype)
c
dm_fuel_con(1,1,1)=1000.0
dm_fuel_con(1,1,2)=1000.0
call fuel_aggr(dm_fuel_con,dmfc_allyr)
ds_fuel_con(1,1,1)=2000.0
ds_fuel_con(1,1,2)=2000.0
call fuel_aggr(ds_fuel_con,dsfc_allyr)
write(*,*)'Aggr=',dsfc_allyr(1,1)
return
end
C*******************************************************************************
subroutine fuel_aggr(fuel_con,fc_allyr)
C*******************************************************************************

include 'test.inc'
c
integer*4 v,y,f
real*4 fc_allyr(M_vtype,M_Ftype)
real*8 fuel_con(N_vtype,N_Evlyrs,N_Ftype)
c
do f=1,N_Ftype
do v=1,N_vtype
fc_allyr(v,f)=0.0
do y=1,N_Evlyrs
fc_allyr(v,f)=fc_allyr(v,f)+fuel_con(v,y,f)
if(f.eq.1.and.v.eq.1.and.y.eq.1)
+ write(*,*)fc_allyr(v,f),fuel_con(v,y,f)
end do
end do
end do
return
end

D:contramc7tubaver16opt_test>ftn95 test.for/link [FTN95/Win32 Ver. 4.8.2038 Copyright (C) Salford Software Ltd 1993-2005]
Licensed to: Paul Chu
Organisation: MOTT MACDONALD

NO ERRORS [<TEST> FTN95/Win32 v4.8.2038]
NO ERRORS [<FUEL_CON_OUT> FTN95/Win32 v4.8.2038]
NO ERRORS [<FUEL_AGGR> FTN95/Win32 v4.8.2038]
Creating executable: test.EXE

D:contramc7tubaver16opt_test>test
1000.00 1000.00000000
2000.00 2000.00000000
Aggr= 2000.00

D:contramc7tubaver16opt_test>ftn95 /intl /logl /save /zeroise /INHIBIT_OPT 36 /OPTIMISE test.for /link [FTN95/Win32 Ver. 4.8.2038 Copyright (C) Salford Software Ltd 1993-2005]
Licensed to: Paul Chu
Organisation: MOTT MACDONALD

NO ERRORS [<TEST> FTN95/Win32 v4.8.2038]
NO ERRORS [<FUEL_CON_OUT> FTN95/Win32 v4.8.2038]
NO ERRORS [<FUEL_AGGR> FTN95/Win32 v4.8.2038]
Creating executable: test.EXE

D:contramc7tubaver16opt_test>test
0.00000 1000.00000000
0.00000 2000.00000000
Aggr= 0.00000

------------------------------------------------------------

The test program works fine without the OPTIMISE switch. However, with the /OPTIMISE switch it produces wrong results.


Ananda

[PaulLaidler]

Ananda

Your program reveals a bug the optimiser that is affecting the line

fc_allyr(v,f)=fc_allyr(v,f)+fuel_con(v,y,f)

The problem is caused by the mixed precision (*4 and * in the expression.

There are various ways around this (until the bug is fixed)

1) Change all REAL*4 to REAL*8 or
2) Break the assignment into two parts to avoid adding REAL*4 to REAL*8
3) Use /INHIBIT_OPT 30 on this part of the code.

Note for other users reading this correspondance...
1) There may not be much improvement in performance when using /OPTIMISE except perhaps during intensive numerical calculations. If you want do use /OPTIMISE, do some timing tests to see if there is a significant improvement in the runtime. If there is no improvement, then for safety, do not use /OPTIMISE.

2) /INHIBIT_OPT will not generally be available until the next release.

[Anonymous]

Paul,

Thank you for the solution to this problem. I had a test with and without /OPTIMISE swith and run showed about 2 minutes saving on a 11 minute run. That is why I wanted to use the /OPTIMISE switch.

I prefer to convert all the code to REAL*8 and continue.

Regards
Ananda

Ananda

[PaulLaidler]

This bug has now been fixed.