1 |
% $Header$ |
% $Header$ |
2 |
% $Name$ |
% $Name$ |
3 |
|
|
4 |
|
Author: Patrick Heimbach |
5 |
|
|
6 |
{\sf Automatic differentiation} (AD), also referred to as algorithmic |
{\sf Automatic differentiation} (AD), also referred to as algorithmic |
7 |
(or, more loosely, computational) differentiation, involves |
(or, more loosely, computational) differentiation, involves |
8 |
automatically deriving code to calculate |
automatically deriving code to calculate |
44 |
%********************************************************************** |
%********************************************************************** |
45 |
\section{Some basic algebra} |
\section{Some basic algebra} |
46 |
\label{sec_ad_algebra} |
\label{sec_ad_algebra} |
47 |
|
\begin{rawhtml} |
48 |
|
<!-- CMIREDIR:sec_ad_algebra: --> |
49 |
|
\end{rawhtml} |
50 |
%********************************************************************** |
%********************************************************************** |
51 |
|
|
52 |
Let $ \cal{M} $ be a general nonlinear, model, i.e. a |
Let $ \cal{M} $ be a general nonlinear, model, i.e. a |
681 |
%********************************************************************** |
%********************************************************************** |
682 |
\section{TLM and ADM generation in general} |
\section{TLM and ADM generation in general} |
683 |
\label{sec_ad_setup_gen} |
\label{sec_ad_setup_gen} |
684 |
|
\begin{rawhtml} |
685 |
|
<!-- CMIREDIR:sec_ad_setup_gen: --> |
686 |
|
\end{rawhtml} |
687 |
%********************************************************************** |
%********************************************************************** |
688 |
|
|
689 |
In this section we describe in a general fashion |
In this section we describe in a general fashion |
690 |
the parts of the code that are relevant for automatic |
the parts of the code that are relevant for automatic |
691 |
differentiation using the software tool TAMC. |
differentiation using the software tool TAF. |
692 |
|
|
693 |
\input{part5/doc_ad_the_model} |
\input{part5/doc_ad_the_model} |
694 |
|
|
695 |
The basic flow is depicted in \ref{fig:adthemodel}. |
The basic flow is depicted in \ref{fig:adthemodel}. |
696 |
If the option {\tt ALLOW\_AUTODIFF\_TAMC} is defined, the driver routine |
If CPP option {\tt ALLOW\_AUTODIFF\_TAMC} is defined, the driver routine |
697 |
{\it the\_model\_main}, instead of calling {\it the\_main\_loop}, |
{\it the\_model\_main}, instead of calling {\it the\_main\_loop}, |
698 |
invokes the adjoint of this routine, {\it adthe\_main\_loop}, |
invokes the adjoint of this routine, {\it adthe\_main\_loop}, |
699 |
which is the toplevel routine in terms of reverse mode computation. |
which is the toplevel routine in terms of automatic differentiation. |
700 |
The routine {\it adthe\_main\_loop} has been generated by TAMC. |
The routine {\it adthe\_main\_loop} has been generated by TAF. |
701 |
It contains both the forward integration of the full model, |
It contains both the forward integration of the full model, the |
702 |
|
cost function calculation, |
703 |
any additional storing that is required for efficient checkpointing, |
any additional storing that is required for efficient checkpointing, |
704 |
and the reverse integration of the adjoint model. |
and the reverse integration of the adjoint model. |
705 |
The structure of {\it adthe\_main\_loop} has been strongly |
|
706 |
simplified for clarification; in particular, no checkpointing |
[DESCRIBE IN A SEPARATE SECTION THE WORKING OF THE TLM] |
707 |
|
|
708 |
|
In Fig. \ref{fig:adthemodel} |
709 |
|
the structure of {\it adthe\_main\_loop} has been strongly |
710 |
|
simplified to focus on the essentials; in particular, no checkpointing |
711 |
procedures are shown here. |
procedures are shown here. |
712 |
Prior to the call of {\it adthe\_main\_loop}, the routine |
Prior to the call of {\it adthe\_main\_loop}, the routine |
713 |
{\it ctrl\_unpack} is invoked to unpack the control vector |
{\it ctrl\_unpack} is invoked to unpack the control vector |
722 |
the gradient has been computed via the adjoint |
the gradient has been computed via the adjoint |
723 |
(cf. Section \ref{section_grdchk}). |
(cf. Section \ref{section_grdchk}). |
724 |
|
|
725 |
|
%------------------------------------------------------------------ |
726 |
|
|
727 |
|
\subsection{General setup |
728 |
|
\label{section_ad_setup}} |
729 |
|
|
730 |
|
In order to configure AD-related setups the following packages need |
731 |
|
to be enabled: |
732 |
|
{\it |
733 |
|
\begin{table}[h!] |
734 |
|
\begin{tabular}{l} |
735 |
|
autodiff \\ |
736 |
|
ctrl \\ |
737 |
|
cost \\ |
738 |
|
grdchk \\ |
739 |
|
\end{tabular} |
740 |
|
\end{table} |
741 |
|
} |
742 |
|
The packages are enabled by adding them to your experiment-specific |
743 |
|
configuration file |
744 |
|
{\it packages.conf} (see Section ???). |
745 |
|
|
746 |
|
The following AD-specific CPP option files need to be customized: |
747 |
|
% |
748 |
|
\begin{itemize} |
749 |
|
% |
750 |
|
\item {\it ECCO\_CPPOPTIONS.h} \\ |
751 |
|
This header file collects CPP options for the packages |
752 |
|
{\it autodiff, cost, ctrl} as well as AD-unrelated options for |
753 |
|
the external forcing package {\it exf}. |
754 |
|
\footnote{NOTE: These options are not set in their package-specific |
755 |
|
headers such as {\it COST\_CPPOPTIONS.h}, but are instead collected |
756 |
|
in the single header file {\it ECCO\_CPPOPTIONS.h}. |
757 |
|
The package-specific header files serve as simple |
758 |
|
placeholders at this point.} |
759 |
|
% |
760 |
|
\item {\it tamc.h} \\ |
761 |
|
This header configures the splitting of the time stepping loop |
762 |
|
w.r.t. the 3-level checkpointing (see section ???). |
763 |
|
|
764 |
|
% |
765 |
|
\end{itemize} |
766 |
|
|
767 |
|
%------------------------------------------------------------------ |
768 |
|
|
769 |
|
\subsection{Building the AD code |
770 |
|
\label{section_ad_build}} |
771 |
|
|
772 |
|
The build process of an AD code is very similar to building |
773 |
|
the forward model. However, depending on which AD code one wishes |
774 |
|
to generate, and on which AD tool is available (TAF or TAMC), |
775 |
|
the following {\tt make} targets are available: |
776 |
|
|
777 |
|
\begin{table}[h!] |
778 |
|
{\footnotesize |
779 |
|
\begin{tabular}{ccll} |
780 |
|
~ & {\it AD-target} & {\it output} & {\it description} \\ |
781 |
|
\hline |
782 |
|
\hline |
783 |
|
(1) & {\tt <MODE><TOOL>only} & {\tt <MODE>\_<TOOL>\_output.f} & |
784 |
|
generates code for $<$MODE$>$ using $<$TOOL$>$ \\ |
785 |
|
~ & ~ & ~ & no {\tt make} dependencies on {\tt .F .h} \\ |
786 |
|
~ & ~ & ~ & useful for compiling on remote platforms \\ |
787 |
|
\hline |
788 |
|
(2) & {\tt <MODE><TOOL>} & {\tt <MODE>\_<TOOL>\_output.f} & |
789 |
|
generates code for $<$MODE$>$ using $<$TOOL$>$ \\ |
790 |
|
~ & ~ & ~ & includes {\tt make} dependencies on {\tt .F .h} \\ |
791 |
|
~ & ~ & ~ & i.e. input for $<$TOOL$>$ may be re-generated \\ |
792 |
|
\hline |
793 |
|
(3) & {\tt <MODE>all} & {\tt mitgcmuv\_<MODE>} & |
794 |
|
generates code for $<$MODE$>$ using $<$TOOL$>$ \\ |
795 |
|
~ & ~ & ~ & and compiles all code \\ |
796 |
|
~ & ~ & ~ & (use of TAF is set as default) \\ |
797 |
|
\hline |
798 |
|
\hline |
799 |
|
\end{tabular} |
800 |
|
} |
801 |
|
\end{table} |
802 |
|
% |
803 |
|
Here, the following placeholders are used |
804 |
|
% |
805 |
|
\begin{itemize} |
806 |
|
% |
807 |
|
\item [$<$TOOL$>$] |
808 |
|
% |
809 |
|
\begin{itemize} |
810 |
|
% |
811 |
|
\item {\tt TAF} |
812 |
|
\item {\tt TAMC} |
813 |
|
% |
814 |
|
\end{itemize} |
815 |
|
% |
816 |
|
\item [$<$MODE$>$] |
817 |
|
% |
818 |
|
\begin{itemize} |
819 |
|
% |
820 |
|
\item {\tt ad} generates the adjoint model (ADM) |
821 |
|
\item {\tt ftl} generates the tangent linear model (TLM) |
822 |
|
\item {\tt svd} generates both ADM and TLM for \\ |
823 |
|
singular value decomposition (SVD) type calculations |
824 |
|
% |
825 |
|
\end{itemize} |
826 |
|
% |
827 |
|
\end{itemize} |
828 |
|
|
829 |
|
For example, to generate the adjoint model using TAF after routines ({\tt .F}) |
830 |
|
or headers ({\tt .h}) have been modified, but without compilation, |
831 |
|
type {\tt make adtaf}; |
832 |
|
or, to generate the tangent linear model using TAMC without |
833 |
|
re-generating the input code, type {\tt make ftltamconly}. |
834 |
|
|
835 |
|
|
836 |
|
A typical full build process to generate the ADM via TAF would |
837 |
|
look like follows: |
838 |
|
\begin{verbatim} |
839 |
|
% mkdir build |
840 |
|
% cd build |
841 |
|
% ../../../tools/genmake2 -mods=../code_ad |
842 |
|
% make depend |
843 |
|
% make adall |
844 |
|
\end{verbatim} |
845 |
|
|
846 |
|
%------------------------------------------------------------------ |
847 |
|
|
848 |
|
\subsection{The AD build process in detail |
849 |
|
\label{section_ad_build_detail}} |
850 |
|
|
851 |
|
The {\tt make <MODE>all} target consists of the following procedures: |
852 |
|
|
853 |
|
\begin{enumerate} |
854 |
|
% |
855 |
|
\item |
856 |
|
A header file {\tt AD\_CONFIG.h} is generated which contains a CPP option |
857 |
|
on which code ought to be generated. Depending on the {\tt make} target, |
858 |
|
the contents is |
859 |
|
\begin{itemize} |
860 |
|
\item |
861 |
|
{\tt \#define ALLOW\_ADJOINT\_RUN} |
862 |
|
\item |
863 |
|
{\tt \#define ALLOW\_TANGENTLINEAR\_RUN} |
864 |
|
\item |
865 |
|
{\tt \#define ALLOW\_ECCO\_OPTIMIZATION} |
866 |
|
\end{itemize} |
867 |
|
% |
868 |
|
\item |
869 |
|
A single file {\tt <MODE>\_input\_code.f} is concatenated |
870 |
|
consisting of all {\tt .f} files that are part of the list {\bf AD\_FILES} |
871 |
|
and all {\tt .flow} files that are part of the list {\bf AD\_FLOW\_FILES}. |
872 |
|
% |
873 |
|
\item |
874 |
|
The AD tool is invoked with the {\bf <MODE>\_<TOOL>\_FLAGS}. |
875 |
|
The default AD tool flags in {\tt genmake2} can be overrwritten by |
876 |
|
an {\tt adjoint\_options} file (similar to the platform-specific |
877 |
|
{\tt build\_options}, see Section ???. |
878 |
|
The AD tool writes the resulting AD code into the file |
879 |
|
{\tt <MODE>\_input\_code\_ad.f} |
880 |
|
% |
881 |
|
\item |
882 |
|
A short sed script {\tt adjoint\_sed} is applied to |
883 |
|
{\tt <MODE>\_input\_code\_ad.f} |
884 |
|
to reinstate {\bf myThid} into the CALL argument list of active file I/O. |
885 |
|
The result is written to file {\tt <MODE>\_<TOOL>\_output.f}. |
886 |
|
% |
887 |
|
\item |
888 |
|
All routines are compiled and an executable is generated |
889 |
|
(see Table ???). |
890 |
|
% |
891 |
|
\end{enumerate} |
892 |
|
|
893 |
|
\subsubsection{The list AD\_FILES and {\tt .list} files} |
894 |
|
|
895 |
|
Not all routines are presented to the AD tool. |
896 |
|
Routines typically hidden are diagnostics routines which |
897 |
|
do not influence the cost function, but may create |
898 |
|
artificial flow dependencies such as I/O of active variables. |
899 |
|
|
900 |
|
{\tt genmake2} generates a list (or variable) {\bf AD\_FILES} |
901 |
|
which contains all routines that are shown to the AD tool. |
902 |
|
This list is put together from all files with suffix {\tt .list} |
903 |
|
that {\tt genmake2} finds in its search directories. |
904 |
|
The list file for the core MITgcm routines is in {\tt model/src/} |
905 |
|
is called {\tt model\_ad\_diff.list}. |
906 |
|
Note that no wrapper routine is shown to TAF. These are either |
907 |
|
not visible at all to the AD code, or hand-written AD code |
908 |
|
is available (see next section). |
909 |
|
|
910 |
|
Each package directory contains its package-specific |
911 |
|
list file {\tt <PKG>\_ad\_diff.list}. For example, |
912 |
|
{\tt pkg/ptracers/} contains the file {\tt ptracers\_ad\_diff.list}. |
913 |
|
Thus, enabling a package will automatically extend the |
914 |
|
{\bf AD\_FILES} list of {\tt genmake2} to incorporate the |
915 |
|
package-specific routines. |
916 |
|
Note that you will need to regenerate the {\tt Makefile} if |
917 |
|
you enable a package (e.g. by adding it to {\tt packages.conf}) |
918 |
|
and a {\tt Makefile} already exists. |
919 |
|
|
920 |
|
\subsubsection{The list AD\_FLOW\_FILES and {\tt .flow} files} |
921 |
|
|
922 |
|
TAMC and TAF can evaluate user-specified directives |
923 |
|
that start with a specific syntax ({\tt CADJ}, {\tt C\$TAF}, {\tt !\$TAF}). |
924 |
|
The main categories of directives are STORE directives and |
925 |
|
FLOW directives. Here, we are concerned with flow directives, |
926 |
|
store directives are treated elsewhere. |
927 |
|
|
928 |
|
Flow directives enable the AD tool to evaluate how it should treat |
929 |
|
routines that are 'hidden' by the user, i.e. routines which are |
930 |
|
not contained in the {\bf AD\_FILES} list (see previous section), |
931 |
|
but which are called in part of the code that the AD tool does see. |
932 |
|
The flow directive tell the AD tool |
933 |
|
% |
934 |
|
\begin{itemize} |
935 |
|
% |
936 |
|
\item which subroutine arguments are input/output |
937 |
|
\item which subroutine arguments are active |
938 |
|
\item which subroutine arguments are required to compute the cost |
939 |
|
\item which subroutine arguments are dependent |
940 |
|
% |
941 |
|
\end{itemize} |
942 |
|
% |
943 |
|
The syntax for the flow directives can be found in the |
944 |
|
AD tool manuals. |
945 |
|
|
946 |
|
{\tt genmake2} generates a list (or variable) {\bf AD\_FLOW\_FILES} |
947 |
|
which contains all files with suffix{\tt .flow} that it finds |
948 |
|
in its search directories. |
949 |
|
The flow directives for the core MITgcm routines of |
950 |
|
{\tt eesupp/src/} and {\tt model/src/} |
951 |
|
reside in {\tt pkg/autodiff/}. |
952 |
|
This directory also contains hand-written adjoint code |
953 |
|
for the MITgcm WRAPPER (see Section ???). |
954 |
|
|
955 |
|
Flow directives for package-specific routines are contained in |
956 |
|
the corresponding package directories in the file |
957 |
|
{\tt <PKG>\_ad.flow}, e.g. ptracers-specific directives are in |
958 |
|
{\tt ptracers\_ad.flow}. |
959 |
|
|
960 |
|
\subsubsection{Store directives for 3-level checkpointing} |
961 |
|
|
962 |
|
The storing that is required at each period of the |
963 |
|
3-level checkpointing is controled by three |
964 |
|
top-level headers. |
965 |
|
|
966 |
|
\begin{verbatim} |
967 |
|
do ilev_3 = 1, nchklev_3 |
968 |
|
# include ``checkpoint_lev3.h'' |
969 |
|
do ilev_2 = 1, nchklev_2 |
970 |
|
# include ``checkpoint_lev2.h'' |
971 |
|
do ilev_1 = 1, nchklev_1 |
972 |
|
# include ``checkpoint_lev1.h'' |
973 |
|
|
974 |
|
... |
975 |
|
|
976 |
|
end do |
977 |
|
end do |
978 |
|
end do |
979 |
|
\end{verbatim} |
980 |
|
|
981 |
|
All files {\tt checkpoint\_lev?.h} are contained in directory |
982 |
|
{\tt pkg/autodiff/}. |
983 |
|
|
984 |
|
|
985 |
|
\subsubsection{Changing the default AD tool flags: ad\_options files} |
986 |
|
|
987 |
|
|
988 |
|
\subsubsection{Hand-written adjoint code} |
989 |
|
|
990 |
|
%------------------------------------------------------------------ |
991 |
|
|
992 |
\subsection{The cost function (dependent variable) |
\subsection{The cost function (dependent variable) |
993 |
\label{section_cost}} |
\label{section_cost}} |
994 |
|
|
1006 |
|
|
1007 |
\input{part5/doc_cost_flow} |
\input{part5/doc_cost_flow} |
1008 |
|
|
1009 |
\subsubsection{genmake and CPP options} |
\subsubsection{Enabling the package} |
1010 |
% |
|
|
\begin{itemize} |
|
|
% |
|
|
\item |
|
1011 |
\fbox{ |
\fbox{ |
1012 |
\begin{minipage}{12cm} |
\begin{minipage}{12cm} |
1013 |
{\it genmake}, {\it CPP\_OPTIONS.h}, {\it ECCO\_CPPOPTIONS.h} |
{\it packages.conf}, {\it ECCO\_CPPOPTIONS.h} |
1014 |
\end{minipage} |
\end{minipage} |
1015 |
} |
} |
1016 |
\end{itemize} |
\begin{itemize} |
|
% |
|
|
The directory {\it pkg/cost} can be included to the |
|
|
compile list in 3 different ways (cf. Section \ref{???}): |
|
1017 |
% |
% |
1018 |
\begin{enumerate} |
\item |
1019 |
|
The package is enabled by adding {\it cost} to your file {\it packages.conf} |
1020 |
|
(see Section ???) |
1021 |
% |
% |
1022 |
\item {\it genmake}: \\ |
\item |
1023 |
Change the default settings in the file {\it genmake} by adding |
|
1024 |
{\bf cost} to the {\bf enable} list (not recommended). |
|
1025 |
% |
\end{itemize} |
|
\item {\it .genmakerc}: \\ |
|
|
Customize the settings of {\bf enable}, {\bf disable} which are |
|
|
appropriate for your experiment in the file {\it .genmakerc} |
|
|
and add the file to your compile directory. |
|
|
% |
|
|
\item genmake-options: \\ |
|
|
Call {\it genmake} with the option |
|
|
{\tt genmake -enable=cost}. |
|
1026 |
% |
% |
1027 |
\end{enumerate} |
|
1028 |
N.B.: In general the following packages ought to be enabled |
N.B.: In general the following packages ought to be enabled |
1029 |
simultaneously: {\it autodiff, cost, ctrl}. |
simultaneously: {\it autodiff, cost, ctrl}. |
1030 |
The basic CPP option to enable the cost function is {\bf ALLOW\_COST}. |
The basic CPP option to enable the cost function is {\bf ALLOW\_COST}. |