odepack

module ODEPACK

This is a WIP(Work In Progress) to evaluate the effort required and feasibility of updating older Fortran code from the netlib repository using a combination of the commercial plusFORT/spag software and GNU utilities as well as conventional manual refactoring.

Many thanks to plusFORT for making an evaluation copy available for several months to the Fortran Community. The plusFORT tools were crucial to this study and unmatched in my experience for removing deprecated syntax from pre-f2003 code.

This began with ODEPACK

the ODEPACK package has been selected as a significant code that is well documented and structured for its vintage, and available on the netlib site and covered by a public domain license.

preliminary targets for phase I:

remove obsolescent syntax (trying plusFORT and spag)
able to build using fpm(1) (the Fortran Package Manager)
in debug mode (ongoing)
with ifort, gfortran, nvfortran
text viewable in ford(1) and extractable as markup that can run through pandoc(1) (ongoing)
available on github (or equivalent)
no common blocks
no equivalences
build as a (single?) module M_odepack (ongoing)
complete set of unit tests (tests currently only use the original examples)

The originally bundled subset of BLAS/LAPACK routines are being included in the module. In production, this might not be done in order to be able to easily call external optimized versions.

The biggest hinderance is some storage used for both INTEGER and DOUBLE PRECISION values.

One take-away is how critical unit tests are to enable rapid development (which so far this package does not have)

The initial pass was done just using the original sample programs as unit tests. This may have allowed for introduction of errors as this is a WIP but the original samples run with the same output as the original.

plusFORT was invaluable and reduced the effort by an estimated 85 percent. The results have been encouraging enough to inspire completing the transformation.

Phase II ?

Some of the goals of phase I are incomplete, but The results of this first pass were significant enough that this project will hopefully continue.

A complete unit test suite is required to allow development to proceed rapidily. Contributions, especially from current ODEPACK users are particularly welcome.

Another major issue is the remaining non-standard code. Non-standard (but at the time de-facto fortran standard) such as equivalencing different types, creating scratch space that is used as different numeric types, and treating scalars as arrays and vice-versa as well as passing the same arrays or values multiple times are the most time-consuming usages to correct to standard-conforming, particularly since spag(1) had already done an excellent job with updating the pre-f2003 code. spag(1) is not (currently?) sufficient by itself to automate the additional refactoring desired, which includes using post-f95 features and code restructuring, so the remaining work requires manual recoding.

The type-mismatch issues have not been eliminated enough to include all the routines in the module, so those in the files “M_da1/dprep.inc” “M_da1/dainvgs.inc” “M_da1/dprepi.inc” “M_da1/dstodi.inc” and “M_da1_/dstode.inc” still require being built without an interface definition.

building with fpm(1)

This version of ODEPACK already builds with an included make(1) file and as an fpm(1) package with the current options:

 fpm run                     --compiler nvfortran --example '*'
 fpm run --profile release   --compiler ifort     --example '*'
 # gfortran for production
 fpm run --profile release --flag -fallow-argument-mismatch --compiler gfortran  --example '*'
 # gfortran for debug
 fpm run --profile debug --flag -fallow-argument-mismatch --flag -std=f2018 --compiler gfortran  --example '*' --verbose

building with make(1)

cd src
# gfortran
make clean
make gfortran
make run
make test

# ifort
make clean
make ifort
make run
make test

# nvfortran
make clean
make nvfortran
make run
make test

To rebuild the documentation with ford(1), use

 ford ford.md

The code is far more readable after having been refactored by a combination of using spag(1) from the plusFORT package and manual editing, and is believed as useable as the original.

There are a few notes in src/M_odepack.f90 concerning continuing issues.

Current users of ODEPACK are encouraged to try this version and provide feedback.

Hopefully as a community we can complete creating a new maintained production-quality version of this venerable and still-valuable package.

Documentation

The ongoing API documentation for the current master branch can be found here. This is generated by processing the source files with FORD.

Solvers for explicitly given systems.

Links to the solver-specific documenation for the main procedures (as described below):

Here is an overview primarily from the original documentation …

Brief Description of ODEPACK - A Systematized Collection of ODE Solvers (Double Precision Version)

Alan C. Hindmarsh
Center for Applied Scientific Computing, L-561
Lawrence Livermore National Laboratory
Livermore, CA 94551, U.S.A.

20 June 2001

Work performed under the auspices of the U.S. Department of Energy by the Lawrence Livermore National Laboratory under contract No. W-7405-Eng-48, and supported (formerly) by the DOE Office of Energy Research, Applied Mathematical Sciences Research Program.

ODEPACK is a collection of Fortran solvers for the initial value problem for ordinary differential equation systems. It consists of nine solvers, namely a basic solver called LSODE and eight variants of it – LSODES, LSODA, LSODAR, LSODPK, LSODKR, LSODI, LSOIBT, and LSODIS. The collection is suitable for both stiff and nonstiff systems. It includes solvers for systems given in explicit form, dy/dt = f(t,y), and also solvers for systems given in linearly implicit form, A(t,y) dy/dt = g(t,y). Two of the solvers use general sparse matrix solvers for the linear systems that arise. Two others use iterative (preconditioned Krylov) methods instead of direct methods for these linear systems. The most recent addition is LSODIS, which solves implicit problems with general sparse treatment of all matrices involved.

The ODEPACK solvers are written in standard Fortran 77, with a few exceptions, and with minimal machine dependencies. There are separate double and single precision versions of ODEPACK. The actual solver names are those given above with a prefix of D- or S- for the double or single precision version, respectively, i.e. DLSODE/SLSODE, etc. Each solver consists of a main driver subroutine having the same name as the solver and some number of subordinate routines. For each solver, there is also a demonstration program, which solves one or two simple problems in a somewhat self-checking manner.

Recently, the ODEPACK solvers were upgraded to improve their portability in numerous ways. Among the improvements are (a) renaming of routines and Common blocks to distinguish double and single precision versions, (b) use of generic intrinsic function names, (c) elimination of the Block Data subprogram, (d) use of a portable routine to set the unit roundoff, and (e) passing of quoted strings to the error message handler. In addition, the prologue and internal comments were reformatted, and use mixed upper/lower case. Numerous minor corrections and improvements were also made.

The above upgrade operations were applied to LSODE earlier than they were to the rest of ODEPACK, and the two upgrades were done somewhat independently. As a result, some differences will be apparent in the source files of LSODE and the other solvers – primarily in the formatting of the comment line prologue of the main driver routine. In Subroutines DLSODE/SLSODE and their subordinate routines, the prologue was written in “SLATEC format”, while for the other solvers a more relaxed style was used. The differences are entirely cosmetic, however, and do not affect performance.

Documentation on the usage of each solver is provided in the initial block of comment lines in the source file, which (in most cases) includes a simple example. A demonstration program (in separate double/single precision versions) is also available.

What follows is a summary of the capabilities of ODEPACK, comments about usage documentation, and notes about installing the collection. For additional documentation on ODEPACK, see also the papers [1], [2] (for LSODE), and [3] (for LSODPK and LSODKR), and in the references cited there. (However, the document [2] does not reflect the upgrade operations described above.)

References:

A. C. Hindmarsh, “ODEPACK, A Systematized Collection of ODE Solvers,” in Scientific Computing, R. S. Stepleman et al. (eds.), North-Holland, Amsterdam, 1983 (vol. 1 of IMACS Transactions on Scientific Computation), pp. 55-64.
K. Radhakrishnan and A. C. Hindmarsh, “Description and Use of LSODE, the Livermore Solver for Ordinary Differential Equations,” LLNL report UCRL-ID-113855, December 1993.
P. N. Brown and A. C. Hindmarsh, “Reduced Storage Matrix Methods in Stiff ODE Systems,” J. Appl. Math. & Comp., 31 (1989), pp.40-91.

I. Summary of the ODEPACK Solvers

A. Solvers for explicitly given systems.

For each of the following solvers, it is assumed that the ODEs are given explicitly, so that the system can be written in the form dy/dt = f(t,y), where y is the vector of dependent variables, and t is the independent variable.

LSODE (Livermore Solver for Ordinary Differential Equations) is the basic solver of the collection. It solves stiff and nonstiff systems of the form dy/dt = f. In the stiff case, it treats the Jacobian matrix df/dy as either a dense (full) or a banded matrix, and as either user-supplied or internally approximated by difference quotients. It uses Adams methods (predictor-corrector) in the nonstiff case, and Backward Differentiation Formula (BDF) methods (the Gear methods) in the stiff case. The linear systems that arise are solved by direct methods (LU factor/solve). LSODE supersedes the older GEAR and GEARB packages, and reflects a complete redesign of the user interface and internal organization, with some algorithmic improvements.
LSODES, written jointly with A. H. Sherman, solves systems dy/dt = f and in the stiff case treats the Jacobian matrix in general sparse form. It determines the sparsity structure on its own, or optionally accepts this information from the user. It then uses parts of the Yale Sparse Matrix Package (YSMP) to solve the linear systems that arise, by a sparse (direct) LU factorization/backsolve method. LSODES supersedes, and improves upon, the older GEARS package.
LSODA, written jointly with L. R. Petzold, solves systems dy/dt = f with a dense or banded Jacobian when the problem is stiff, but it automatically selects between nonstiff (Adams) and stiff (BDF) methods. It uses the nonstiff method initially, and dynamically monitors data in order to decide which method to use.
LSODAR, also written jointly with L. R. Petzold, is a variant of LSODA with a rootfinding capability added. Thus it solves problems dy/dt = f with dense or banded Jacobian and automatic method selection, and at the same time, it finds the roots of any of a set of given functions of the form g(t,y). This is often useful for finding stop conditions, or for finding points at which a switch is to be made in the function f.
LSODPK, written jointly with Peter N. Brown, is a variant of LSODE in which the direct solvers for the linear systems have been replaced by a selection of four preconditioned Krylov (iterative) solvers. The user must supply a pair of routine to evaluate, preprocess, and solve the (left and/or right) preconditioner matrices. LSODPK also includes an option for a user-supplied linear system solver to be used without Krylov iteration.
LSODKR is a variant of LSODPK with the addition of the same rootfinding capability as in LSODAR, and also of automatic switching between functional and Newton iteration. The nonlinear iteration method-switching differs from the method-switching in LSODA and LSODAR, but provides similar savings by using the cheaper method in the non-stiff regions of the problem. LSODKR also improves on the Krylov methods in LSODPK by offering the option to save and reuse the approximate Jacobian data underlying the preconditioner.

B. Solvers for linearly implicit systems.

The following solvers treat systems in the linearly implicit form A(t,y) dy/dt = g(t,y), A = a square matrix, i.e. with the derivative dy/dt implicit, but linearly so. These solvers allow A to be singular, in which case the system is a differential-algebraic equation (DAE) system. In that case, the user must be very careful to supply a well-posed problem with consistent initial conditions.

LSODI, written jointly with J. F. Painter, solves linearly implicit systems in which the matrices involved (A, dg/dy, and d(A dy/dt)/dy) are all assumed to be either dense or banded. LSODI supersedes the older GEARIB solver and improves upon it in numerous ways.
LSOIBT, written jointly with C. S. Kenney, solves linearly implicit systems in which the matrices involved are all assumed to be block-tridiagonal. Linear systems are solved by the LU method.
LSODIS, written jointly with S. Balsdon, solves linearly implicit systems in which the matrices involved are all assumed to be sparse. Like LSODES, LSODIS either determines the sparsity structure or accepts it from the user, and uses parts of the Yale Sparse Matrix Package to solve the linear systems that arise, by a direct method.

II. Usage Documentation

Each of the solvers in the ODEPACK collection is headed by a user-callable driver subroutine, with the same name as the solver (DLSODE, etc.). The call sequence of the driver routine includes the names of one or more user-supplied subroutines that define the ODE system, and various other problem and solution parameters. Complete user documentation is given in the initial block of comment lines (the prologue) in the driver routine. In each case, this prologue is organized as follows:

Summary of Usage (short, for standard modes of use)
Example Problem, with code and output (except for LSODPK and LSODKR)
Full Description of User Interface, further divided as follows:
1. Call sequence description (including optional inputs/outputs)
2. Optionally callable routines
3. Descriptions of internal Common blocks
4. Optionally user-replaceable routines
Revision History, showing date written and dates of revisions
Other Routines, a list of all subordinate routines for the solver

First-time users should read only the Summary of Usage and look at the the Example Problem (or demonstration program), then later refer to the Full Description if and when more details or nonstandard options are needed.

III. Installation Notes

Use of supplied matrix procedures

THe src/M_matrix/ directory includes modified versions of routines from the LINPACK and BLAS collections that are needed by the solvers (and by two of the demonstration programs), for the solution of dense and banded linear systems and associated basic linear algebra operations. These routine are:

   _From LINPACK_ :  DGEFA, DGESL, DGBFA, DGBSL
   _From the BLAS_: DAXPY, DCOPY, DDOT, DSCAL, DNRM2, IDAMAX

If your computer system already has these routines, and especially if it has machine-optimized versions, the copies provided in the M_module module should probably not be called by your program if high performance is required.

The message package

The second auxiliary source file directory M_da1/ includes a set of five routines – XERRWD, XSETUN, XSETF, IXSAV, IUMACH – which handle error messages from the solvers.

These routines are slated for elimination and replacement with a more intuitive interface.

This set is in fact a reduced version (sufficient for the needs of ODEPACK) of a much larger error handling package from the SLATEC Library. If your computer system already has the full SLATEC error handler, the version provided here can be ignored. If the reduced version is used, its machine-dependent features should be checked first; see comments in Subroutine XERRWD.

Non-standard code

ODEPACK contains a few instances where the Fortran Standard is violated:

(a) In various places in the LSODES and LSODIS solvers, a call to a subroutine has a subscripted real array as an argument where the subroutine called expects an integer array. Calls of this form occur in Subroutine DLSODES (to DSTODE), in DIPREP (to DPREP), in Subroutine DLSODIS (to DSTODI), and in DIPREPI (to DPREPI). Another such call occurs in the DLSODES demonstration program, from the main program to Subroutine SSOUT.

This is done in order to use work space in an efficient manner, as the same space is sometimes used for real work space and sometimes for integer work space. If your compiler does not accept this feature, one possible way to get the desired result is to compile the called routines and calling routines in separate jobs, and then combine the binary modules in an appropriate manner.

If this procedure is still not acceptable under your system, it will be necessary to radically alter the structure of the array RWORK within the LSODES or LSODIS solver package. (See also Note 5 below.)

(b) Each ODEPACK solver treats the arguments NEQ, Y, RTOL, and ATOL as arrays, even though the length may be only 1. Moreover, except for Y, the usage instructions say that these arguments may be either arrays or scalars. If your system does not allow such a mismatch, then the documentation of these arguments should be changed accordingly.

For maximum storage economy, the LSODES and LSODIS solvers make use of the real to integer wordlength ratio. This is assumed to be an integer L such that if a real array R and an integer array M occupy the same space in memory, R(1) having the same bit address as M(1), then R(I) has the same address as M((I-1)*L+1). This ratio L is usually 2 for double precision, and this is the value used in the double precision version supplied. If this value is incorrect, it needs to be changed in two places:

(a) The integer LENRAT is DATA-loaded in Subroutines DLSODES and DLSODIS to this ratio, shortly below the prologue.

(b) The integer LRATIO is DATA-loaded in Subroutine CDRV to this ratio, shortly below the prologue of that routine.

(See comments in both places.) If the ratio is not an integer, use the greatest integer not exceeding the ratio.

For installation of ODEPACK on a Cray computer, the source files supplied include compiler directives for the CFT compiler. These have the form CDIR$ IVDEP and occur prior to certain loops that involve subscript shifts (and would otherwise not be vectorized). These directives are (or should be) treated as comments by any other compiler.
On first obtaining ODEPACK, the demonstration programs should be compiled and executed prior to any other use of the solvers. In most cases, these exercise all of the major method options in each solver, and are self-checking. (In the case of LSODPK and LSODKR, the demonstration programs are not self-checking, and for LSODKR only one major method option is used.) In any case, the output can be compared with the sample output supplied, which was generated from the double precision version of ODEPACK on a 32-bit computer. When comparing your output with that supplied, differences of 10-20% in the final values of the various statistical counters can be attributed to differences in the roundoff properties of different computer systems.
If some subset of the whole ODEPACK collection is desired, without unneeded routines, the appropriate routines must be extracted accordingly. The following lists give the routines needed for the double precision version of each solver.

The DLSODE solver consists of the routines DLSODE, DINTDY, DSTODE, DCFODE, DPREPJ, DSOLSY, DEWSET, DVNORM, DSRCOM, DGEFA, DGESL, DGBFA, DGBSL, DAXPY, DSCAL, DDOT, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODES solver consists of the routines DLSODES, DIPREP, DPREP, JGROUP, ADJLR, CNTNZU, DINTDY, DSTODE, DCFODE, DPRJS, DSOLSS, DEWSET, DVNORM, DSRCMS, ODRV, MD, MDI, MDM, MDP, MDU, SRO, CDRV, NROC, NSFC, NNFC, NNSC, NNTC, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODA solver consists of the routines DLSODA, DINTDY, DSTODA, DCFODE, DPRJA, DSOLSY, DEWSET, DMNORM, DFNORM, DBNORM, DSRCMA, DGEFA, DGESL, DGBFA, DGBSL, DAXPY, DSCAL, DDOT, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODAR solver consists of the routines DLSODAR, DRCHEK, DROOTS, DINTDY, DSTODA, DCFODE, DPRJA, DSOLSY, DEWSET, DMNORM, DFNORM, DBNORM, DSRCAR, DGEFA, DGESL, DGBFA, DGBSL, DAXPY, DSCAL, DDOT, DCOPY, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODPK solver consists of the routines DLSODPK, DINTDY, DEWSET, DVNORM, DSTODPK, DCFODE, DPKSET, DSOLPK, DSPIOM, DATV, DORTHOG, DHEFA, DHESL, DSPIGMR, DHEQR, DHELS, DPCG, DPCGS, DATP, DUSOL, DSRCPK, DAXPY, DSCAL, DCOPY, DDOT, DNRM2, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODKR solver consists of the routines DLSODKR, DRCHEK, DROOTS, DLHIN, DINTDY, DEWSET, DVNORM, DSTOKA, DCFODE, DSETPK, DSOLPK, DSPIOM, DATV, DORTHOG, DHEFA, DHESL, DSPIGMR, DHEQR, DHELS, DPCG, DPCGS, DATP, DUSOL, DSRCKR, DAXPY, DSCAL, DCOPY, DDOT, DNRM2, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODI solver consists of the routines DLSODI, DAINVG, DINTDY, DSTODI, DCFODE, DPREPJI, DSOLSY, DEWSET, DVNORM, DSRCOM, DGEFA, DGESL, DGBFA, DGBSL, DAXPY, DSCAL, DDOT, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSOIBT solver consists of the routines DLSOIBT, DAIGBT, DINTDY, DSTODI, DCFODE, DPJIBT, DSLSBT, DEWSET, DVNORM, DSRCOM, DDECBT, DSOLBT, DGEFA, DGESL, DAXPY, DSCAL, DDOT, IDAMAX, XERRWD, XSETUN, XSETF, IXSAV, IUMACH

The DLSODIS solver consists of the routines DLSODIS, DAINVGS, DIPREPI, DPREPI, JGROUP, ADJLR, CNTNZU, DINTDY, DSTODI, DCFODE, DPRJIS, DSOLSS, DEWSET, DVNORM, DSRCMS, ODRV, MD, MDI, MDM, MDP, MDU, SRO, CDRV, NROC, NSFC, NNFC, NNSC, NNTC, XERRWD, XSETUN, XSETF, IXSAV, IUMACH