Here are some frequently asked questions (with assorted frequently given answers...). Please check these before contacting us with any problems.
oomph-lib'sinstallation is performed with the gcc compiler suite, using full optimisation and warning (
autogen.shscript you are given the option to overwrite the default compilation flags, either by specifying alternative flags on the command line, or by recycling any of the previously used combinations of flags, stored in the directory config/configure_options. The files in this directory also contain more details on available flags and options.
installinstallation procedure, you should already know how to specify flags at the configure stage...
autotoolsframework, additional flags for the compilation of individual (driver) codes may be specified by setting the appropriate automake variables in the
am. For instance, to use the compiler flag
-DUSE_TAYLOR_HOODduring the compilation of the executable
am. The C++ section of the automake manual provides more detail on other
oomph-libfrom outside the
autotoolsframework, you can, of course, use whatever technique you prefer to customise the compilation of your driver code.
oomph-lib is developed in a GNU Linux environment, using the gcc compiler suite. We believe (!) the source code complies with the C++ standard and it compiles cleanly under the gcc compilers that we have access to (version 3.2.3 and later). Some warning messages tend to be issued during the compilation of the third-party libraries distributed with
oomph-lib, but these are not our responsibility!
We occasionally compile (or get other people to compile) the library under other compilers and we generally try to rectify any problems that are flagged up in the process. If you encounter any problems (errors or warning messages) while compiling the library with your own (non-gcc) compiler we would like to hear from you, especially if you suggest concrete bug fixes.
There are problems with version 7 of the intel compiler suite but the library compiles cleanly under version 9. We suggest you upgrade to that if you wish to use the intel compilers.
Here are a few things to check:
oomph-libinstallation directory. Linux does not allow spaces in directory names!
Here are some general guidelines for installing under OSX, contributed by Rich Hewitt:
As discussed in the installation instructions, Darwin (the BSD-derived UNIX core of Apple's OSX operating system) requires a slight change to the default procedure for linking against
oomph-lib from outside its autotools framework. As a result, the linking self-test tends to fail on machines with this operating system. Instructions on how to fix this problem are provided in installation instructions.
Tolerance of comparison between floating point numbers
oomph-lib distribution contains a large number of demo codes in the sub-directory
demo_drivers. Primarily, these codes serve as demo codes for the
oomph-lib documentation (contained in the sub-directory
doc) but they are also used during the library's self-test procedure which checks that
oomph-lib was installed correctly. The self-tests can either be initiated at the end of the
autogen.sh - based installation or by typing
oomph-lib's top-level directory. The self-test builds and runs all demo codes and compares their results against the reference data stored in the
validata sub-directories. The comparison is performed with the python script
bin/fpdiff.py which tolerates small (machine- and compiler-dependent) roundoff errors and suppresses the comparison of numbers whose absolute value falls below some threshold (to avoid the comparison of numerical zeroes).
If any of the self-test fail, you should first check the output in the file
validation.log to assess if the differences between the computed data and the reference data are significant. If the discrepancy appears to be due to larger-than-anticipated round-off errors (you'll have to judge this yourself!), modify the
validate.sh script to specify a larger relative tolerance and/or a larger value for the threshold below which
fpdiff.py regards numbers as numerical zeroes and excludes them from the comparison. [Type
oomph-lib's top-level directory for instructions on how to use the script].
The self-test fails even though the output files produced by the code are correct
This is an odd error that is usually caused by the use of wildcards in the validations scripts, or the comparison of data that is stored in certain STL containers. Typically, the self-test is performed by the
validate.sh shell script, which runs the executable and concatenates selected output files to a single file whose contents are compared against the reference file in the
While it is tempting to write
it is important to realise that the order in which the files are concatenated is machine- and/or operating-system dependent. If the above command is run in a directory with the following structure
some operating systems will expand the command to
while others will execute
In this case the self-test will report a failure, even though the solution files are correct. The
validate.sh scripts should therefore not contain any wildcards.
Similar problems can arise if the validation data includes data that is stored in certain STL containers such as sets. The order in which items are stored in such containers may vary from machine to machine and from compiler to compiler. If such data is to be included in a self-test the data should be sorted first, based on a user-controllable sorting criterion.
Assume you have followed our advice to explore
oomph-lib by playing with representative demo driver codes, and have ended up in a directory that contains the following files:
The temptation is to compile the driver code
one_d_poisson.cc with your favourite C++ compiler, e.g.
but this does not work:
Why? If you look into the code, you notice that it #includes
generic.h header file, and
g++ obviously doesn't know where to find this. You can, in principle, find out where it (and the associated library files themselves!) live but this is rather against the spirit of the automake machinery. What you should do instead is to use
Note the omitted postfix! The argument to
make is the name of the target, i.e. the executable that you wish to build – which source code to compile, where to locate the required headers and which libraries to link against is all encoded in the
Makefile (which has itself been created by automake (during the installation of the
oomph-lib) from the
Makefile.am file in the directory). As you see from the on-screen output, there is a lot of information that needs to be passed to the compiler/linker:
Looking at the content of the directory again shows that the required executable has now been generated:
If you are familiar with makefiles, you are unlikely to have fallen into the trap discussed above. You will almost certainly have spotted the
Makefile and tried to build the executable(s) with
make, but oddly this doesn't work as you (probably) expected:
The reason for this is that
demo_driversdouble up as self-tests to assert the correct installation of
checkis run from the top-level directory.
The latter requires them to be declared (in the
Makefile.am) in the
check_PROGRAMS variable. You can either build (and automatically run) all executables in a
demo_driver directory by issuing the command
or you can inspect the
check_PROGRAMS variable in the
Makefile.am to find out which targets have been declared (and can therefore be built by
make). For instance, in the
one_d_poisson directory the first few lines of the
This declares the target (the executable called
one_d_poisson) and specifies the source code to be compiled (
one_d_poisson.cc). The build target needs to specified explicitly:
We note that the names of the executable and the source code do not have to be as closely related as in this example. While we usually create the name of the executable by dropping the postfix of the source file, this is not guaranteed! Inspect the
make doesn't give you what you want.
To facilitate the installation, the oomph-lib distribution includes the relevant parts of certain third-party libraries, such as SuperLU and BLAS. These "external libraries" are built and installed along with
oomph-lib's own sub-libraries. To distinguish them from any already existing local versions of these libraries, their names are pre-fixed with the string
All demo codes are automatically linked against these libraries. This is achieved by defining the variable
EXTERNAL_LIBS in the
If you wish to link against some other version of the library, edit the
EXTERNAL_LIBS variable accordingly.
WARNING: If you work within the autotools framework, remember that the
configure script is generated by
autoconf, using the file
configure.ac, which is itself (re-)generated whenever you run
autogen.sh. If you use the autotools, you should edit the
EXTERNAL_LIBS variable in the file
configure/configure.ac_scripts/start which forms one of the building blocks from which
autogen.sh assembles the
configure.ac file. Once you have changed the
EXTERNAL_LIBS variable in
configure/configure.ac_scripts/start you should (re-)run
oomph-lib's top-level directory.
Recent versions of the gcc compilers enforce the C++ standard much more rigorously than earlier versions. Unfortunately, the standard includes some rules that are so counter-intuitive that it is hard get into the habit of using them, especially if code is developed on a compiler that does not enforce the standard as rigorously.
The most frequent problem arises in classes that are derived from a templated base class. The C++ standard insists that all references to member functions (or member data) that is defined in the templated base class must be preceded by "this->" when the reference is made in the derived class. Allegedly, this is necessary to avoid ambiguities, though it is not entirely clear what this ambiguity is supposed to be... Here is a driver code that illustrates the problem.
If you compile this with sufficiently recent versions of the gcc compilers, using the flag
-DUSE_BROKEN_VERSION, the compilation will fail with the following error:
You may not only stumble across this problem in one of your own codes but it is also possible that some code in the library itself still violates this rule. This is because templated classes are only built when needed and it is conceivable that
oomph-lib's suite of self-tests do not instantiate all templated classes that exist in the library. If you encounter any such problems, check if putting a "this->" in front of the function call fixes the problem. If it does, let us know!
When linking, some versions of the gcc compiler produce warnings about references to "discarded sections" being referenced. Here's an example:
We admit to being slightly baffled by this. Other libraries seem to suffer from the same problem (google for .
discarded, say), but as far as we can tell no solution has ever been suggested, nor does one seem to be required. The executable works fine. Upgrade to a newer version of gcc?
Problem::self_test()before solving the problem. This function performs a large number of sanity checks and reports any inconsistencies in the data structure.
oomph-lib'swrapper to the STL vector class, with optional range checking) they can easily be detected. Make sure to re-compile again with RANGE_CHECKING switched off before you start any production runs – the run-time overheads incurred by range-checking are significant!
gdbor its GUI-based equivalent ddd) to (try to!) find out where the segmentation fault occurred. [Careful: If the segmentation fault is caused by a pointer problem, this naive inspection can be quite misleading – tell-tale signs are that the traceback displays a non-sensical call stack, e.g. a function being called "out of nowhere"; variables that have just been given values not existing; etc.]
oomph-lib'sPoisson elements solve the Laplace equation unless a function pointer to the source function is specified.
oomph-lib'stimestepping procedures in the context of the unsteady heat equation for details.
Problem::actions_before_implicit_timestep()to update any time-dependent boundary conditions?
Problem::newton_solve(...) will converge quadratically, provided
oomph-libalso provides automatic continuation methods, based on Keller's arclength continuation, but at the moment, no tutorials exist for these.
oomph-libwill then use the default implementation of these functions in the
GeneralisedElementbase class to compute the Jacobian matrices by finite-differencing. The executable is likely to run more slowly since the finite-difference-based computation is unlikely to be as efficient as the customised implementation for your specific element, but if the Newton method then converges, you know where to look for your bug! You may also want to check for any un-initialised variables. They are the most likely culprits if your code behaves differently at different levels of optimisation as more aggressive optimisation may suppress any default initialisations of data – in fact, you should never rely on that anyway!
SpineNodesor nodes whose position is updated by a
MacroElement/Domain- based procedure), the position of the nodes in the "bulk" mesh must be updated whenever the Newton method updates the unknowns. This is most easily done by calling
oomph-lib's high-level post-processing routines output the results of the computations in a form that is suitable for display with tecplot, a powerful commercial plotting package. Purists may find it odd that an open-source library should choose an output format that is customised for a commercial software package. We tend to agree... Our only excuse is that tecplot is very very good, and without it we would have found it extremely difficult to create many of the plots shown in the tutorials. [If you know of any open-source plotting package whose capabilities are comparable to those of tecplot, let us know!]
Angelo Simone has written a python script that converts
oomph-lib's output to the vtu format that can be read by paraview, an open-source 3D plotting package. The conversion script can currently deal with output from meshes that are composed of 2D quad elements – the extension to 3D is work in progress. Use of the conversion script is documented in another tutorial.
It is possible to display
oomph-lib's default output (in more elementary form, obviously) with gnuplot. The trick is to specify the
using option in gnuplot's plot commands – in this mode gnuplot ignores tecplot's "ZONE" commands. For instance, trying to plot the x-y data created by the demo code for the solution of the 1D Poisson equation with
If the data is too complex to be displayed by gnuplot, you may wish to customise the output for your preferred plotting package. This is easily done as
oomph-lib creates its output element-by-element. The elements' various
output(...) functions are virtual functions that can easily be overloaded in a user-defined wrapper class.
Here is an example driver code that illustrates how to change the output from
QPoissonElement family of 1D-line/2D-quad/3D-brick Poisson elements so that they output the string "Hello world".
poisson library headers:
and then create a customised version of the Poisson elements in which we overload the tecplot-based
QPoissonElement<DIM,NNODE_1D>::output(...) function, defined in the
If we now call the output function, the version defined in the customised element is used. The remaining implementation of the Poisson element remains unchanged.
oomph-lib's equations classes (or elements) are implemented in great generality. For instance, our discretisation of the Navier-Stokes equations includes a source term in the continuity equation, and body force terms in the momentum equations; it allows switching between the stress-divergence and simplified forms of the viscous terms; it includes the mesh velocity into the ALE formulation of the time-derivatives; etc. This makes the elements very versatile and robust. However, the generality/robustness comes at a price: Even though we provide default values for most functions (e.g. the body force terms default to zero), their evaluation requires a finite amount of CPU time. If you wish to use the elements in a simple application in which the Navier-Stokes equations are solved in a fixed domain, without any body forces or other source terms, say, you may wish to disable the additional functionality.
This is easily done: After all,
oomph-lib is open-source software and you can therefore change anything you want! In principle, you could edit the source code in the
src/navier_stokes directory and delete (or at least comment out) all the functionality that you do not require. However, this is probably a risky step as it will break all demo codes (used during
oomph-lib's self-test procedure) that use some of the features that you are not interested in. We therefore recommend copying the content of the directory
src/navier_stokes into a new directory, e.g.
user_src/my_navier_stokes and to edit the copied sources. Follow the instructions on the
oomph-lib installation page to turn these sources into a separate library against which you can link.
The default behaviour when using
check is to compile and run the self tests one test at a time. This means that compilation failures may take a long time to appear which can be frustrating when making changes to the main library. The alternative is to first compile all self tests using
and then once you are sure everything compiles run the self tests as normal.
Note: this "trick" relies on undocumented behaviour in automake and so it is possible (but unlikely) that it will not work in new versions (tested and working with GNU automake 1.11.3).
oomph-lib does contain a lot of code and a lot of documentation. How to get started obviously depends on your background: Are you familiar with the finite element method? How good is your knowledge of C++? Etc.
Here are some possible "routemaps" around the library:
oomph-lib'scapabilities. Pick a problem that interests you and study the associated tutorial. Copy the driver code into your own directory and play with it.
oomph-lib'soverall data structure, or find out how to optimise the library for your particular application.
oomph-libobjects for your problem:
oomph-lib. You should at least understand:
Assume you have studied one of the example codes and wish to find out more about the implementation of a particular class or function that is used there. How do you find its source code and/or its full documentation?
Generally, a class/function that is used in a demo code can only be defined in one of two places:
oomph-lib's tutorials tend to provide a fairly complete annotated listing of the relevant driver codes; if the function you are interested in is not mentioned explicitly in the tutorial, it is most likely to be defined in an include file. You can inspect the driver code in its entirety by following the link at the end of the tutorials. If you cannot find the class/function there, it must be defined in one of the include files listed at the beginning of the source code.
The included files themselves can either be located in the same directory as the demo driver (the directory also tends to be mentioned at the end of the tutorial) or in one of
oomph-lib's sub-libraries. The source code for these is located in the sub-directories of the
src directory. Often the class/function is defined in a source file with an "obvious" name; if not, use
grep to find it. This can, of course, be done recursively. For instance, the command
oomph-lib's top-level directory will search through the entire distribution to locate files that contain the string "FiniteElement".
You can also use the html-based representation of
oomph-lib's data structure, created by doxygen, in the "bottom-up" discussion of the data structure. (Note that the search menu may not work on your browser.)
If all else fails and you think you have found a bug in the library, make sure you follow these steps:
PARANOIDflags are set?
A pdf version of this document is available.