This directory contains code to compile the optimization routine optim.x.
Note, that there is an alternative optimization package optim_m1qn3 that is based on the latest (and last) version of m1qn3.
optim.x requires the library lsopt_ecco and the BLAS library blas. The
build procedure is a two step process (first build lsopt_ecco and blas, and
then optim.x ) and by no means foolproof. It requires that you generate a
makefile in this directory and modify the sample Makefile in ../lsopt. For
illustration purposes, let us assume that we want to generate an optim.x for
myExp=tutorial_global_oce_optim in MITgcm/verification.
In MITgcm/lsopt adjust the compiler and compiler flags in Makefile. Using
the same compiler and flags as for building the mitgcmuv_ad executable in
$myExp is probably the best guess, see also below how the MITgcm/optim
Makefile is generated. The default works for a standard Ubuntu system, but
not e.g. for a Mac. There's a makefile MITgcm/lsopt/Makefile_macos that has
worked for MacOS, but may require adjustment. After adjusting the makefile,
compile the libraries like this:
cd ../lsopt
make
The resulting libraries lsopt_ecco and blas will be used in the second
step.
To generate the makefile based on the setting in $myExp and to build
optim.x, change into optim and run
cd ../optim
./prep_make ../../MITgcm/verification/${myExp}/build
make clean
make depend
make
prep_make generates a local Makefile from makefile_templ and from
$myExp build/Makefile. It fills some makefile_templ placeholder
_GET_keyWord with the corresponding option/parameter value "keyWord" found in
the build/Makefile (the current list of keyWords is: BLD_DIR, EXTRA_OPT,
CPPCMD, SFX, FC, FFLAGS and FOPTIM). A simple usage description is
returned when typing ./prep_make alone.
In some cases you may have to adjust makefile_templ(e.g. for the path to a
non-standard makedepend) before running prep_make.
Note that the Makefile generated using prep_make option -fake is only for
testing purpose, to get a fake optim.x that just reads gradient vector files
without any lsopt pieces.
It may make sense to first generate the makefile in MITgcm/optim with
prep_make, and then use the parameters in the generated Makefile to adjust
the sample Makefile in MITgcm/lsopt, build the libraries, and then
return to MITgcm/optim to build optim.x.
With PR #796, the header format of
the control and gradient vector files and the length of the stored fields has
changed. To read old gradient vector files (typically called
ecco_cost_MIT_CE_000.opt0000 for the first optimisation call) with optim.x,
define CPP flag READ_OLD_CTRL_PACK_FILE in CTRL_OPTIONS.h before compiling
optim.x. The resulting optim.x will read the old format gradient vector and
write the new control vector (typically called ecco_ctrl_MIT_CE_000.opt0001)
in the new format. After the new control vector has been written,
READ_OLD_CTRL_PACK_FILE should be reset to undef again (and optim.x
recompiled) for the next optimisation.
The following is the content of the old README. It describes some sort of
interface, i.e. the header of the control and gradient vectors written and read
by the mitgcmuv_ad (and optim.x), see optim_readdata.F and
optim_writedata.F. More details can be found in the online
manual
(Chapter 10).
c expid - experiment name
c optimcycle - optimization no.
c missing value - missing value identifier (usually -9999.)
c ig - global start index x (zonal)
c jg - global start index y (merid.)
c nsx - no. of x-subgrids
c nsy - no. of y-subgrids
>>> MISSING: <<<
c nr - no. of z-points vertical
c snx
c sny
c nvartype
c nvarlength
>>> <<<
c maxcvars - Number of control variables
c (currently 6; 2 init. + 4 bound.)
integer maxcvars
parameter ( maxcvars = 20 )
c ncvarindex - "arbitrary" index to define variable:
c * 101: initial temp.
c * 102: initial sali.
c * 103: heat flux
c * 104: freshwater flux
c * 105: u stress (zonal)
c * 106: v stress (merid.)
integer ncvarindex ( maxcvars )
c ncvarrecs - no. of records in control vector
c * = 1 for init. temp./sali.
c * = endrec - startrec + 1 for fluxes
integer ncvarrecs ( maxcvars )
c ncvarrecstart - first record:
c * NOT DEFINED for init. temp./sali.
c * = startrec for fluxes
integer ncvarrecstart ( maxcvars )
c ncvarrecsend - last record:
c * NOT DEFINED for init. temp./sali.
c * = endrec for fluxes
integer ncvarrecsend ( maxcvars )
c ncvarxmax - no. of x-points in subgrid (zonal)
c = snx
integer ncvarxmax ( maxcvars )
c ncvarymax - no. of y-points in subgrid (merid.)
c = sny
integer ncvarymax ( maxcvars )
c ncvarnrmax - no. of z-points (vert.)
c * = nr for init. temp./sali.
c * = 1 for fluxes
integer ncvarnrmax ( maxcvars )
c nwet[c/s/w]tile - Number of wet points in a tile for center (c),
c south (s), and western (w) mask, resp.
integer nwetctile ( nsx, nsy, nr )
integer nwetstile ( nsx, nsy, nr )
integer nwetwtile ( nsx, nsy, nr )
c ncvargrd - position in grid
c * = 'c' : (center,center)
c * = 's' : (center,south)
c * = 'w' : (west,center)
character*(1) ncvargrd(maxcvars)