Programmer’s manual

Build system

Yade uses cmake the cross-platform, open-source build system for managing the build process. It takes care of configuration, compilation and installation. CMake is used to control the software compilation process using simple platform and compiler independent configuration files. CMake generates native makefiles and workspaces that can be used in the compiler environment of your choice.


Yade source tree has the following structure (omiting, doc, examples and scripts which don’t participate in the build process); we shall call each top-level component module:

core/         ## core simulation building blocks
extra/        ## miscillanea
gui/          ## user interfaces
   qt4/          ## graphical user interface based on qt3 and OpenGL
   py/           ## python console interface (phased out)
lib/          ## support libraries, not specific to simulations
pkg/          ## simulation-specific files
   common/       ## generally useful classes
   dem/          ## classes for Discrete Element Method
py/           ## python modules

Header installation

To allow flexibility in source layout, CMAKE will copy (symlink) all headers into flattened structure within the build directory. First 2 components of the original directory are joind by dash, deeper levels are discarded (in case of core and extra, only 1 level is used). The following table makes gives a few examples:

Original header location Included as
core/Scene.hpp <core/Scene.hpp>
lib/base/Logging.hpp <lib-base/Logging.hpp>
lib/serialization/Serializable.hpp <lib-serialization/Serializable.hpp>
pkg/dem/DataClass/SpherePack.hpp <pkg-dem/SpherePack.hpp>
gui/qt3/QtGUI.hpp <gui-qt3/QtGUI.hpp>

It is advised to use #include<module/Class.hpp> style of inclusion rather than #include"Class.hpp even if you are in the same directory.

Automatic compilation

In the pkg/ directory, situation is different. In order to maximally ease addition of modules to yade, all *.cpp files are automatically scanned by CMAKE and considered for compilation. Each file may contain multiple lines that declare features that are necessary for this file to be compiled:


This file will be compiled only if both VTK and GTS features are enabled. Depending on current feature set, only selection of plugins will be compiled.

It is possible to disable compilation of a file by requiring any non-existent feature, such as:

YADE_REQUIRE_FEATURE(temporarily disabled 345uiysdijkn);

The YADE_REQUIRE_FEATURE macro expands to nothing during actual compilation.


The order in which modules might depend on each other is given as follows:

module resulting shared library dependencies
lib can depend on external libraries, may not depend on any other part of Yade.
core yade-support; may depend on external libraries.
pkg core, yade-support
gui, lib, core, pkg
py (many files) lib, core, pkg, external

Development tools

Integrated Development Environment and other tools

A frequently used IDE is Kdevelop. We recommend using this software for navigating in the sources, compiling and debugging. Other useful tools for debugging and profiling are Valgrind and KCachegrind. A series of wiki pages is dedicated to these tools in the development section of the wiki.

Hosting and versioning

The Yade project is kindly hosted at launchpad, which is used for source code, bug tracking, planning, package downloads and more.

The versioning software used is GIT, for which a short tutorial can be found in Yade on GitHub. GIT is a distributed revision control system. It is available packaged for all major linux distributions.

The source code is hosted on GitHub , which is periodically imported to Launchpad for building PPA-packages. The repository can be http-browsed.

Build robot

A build robot hosted at 3SR lab. is tracking souce code changes. Each time a change in the source code is commited to the main development branch via GIT, the “buildbot” downloads and compiles the new version, and start a series of tests.

If a compilation error has been introduced, it will be notified to the yade-dev mailing list and to the commiter, thus helping to fix problems quickly. If the compilation is successfull, the buildbot starts unit regression tests and “check tests” (see below) and report the results. If all tests are passed, a new version of the documentation is generated and uploaded to the website in html and pdf formats. As a consequence, those two links always point to the documentation (the one you are reading now) of the last successfull build, and the delay between commits and documentation updates are very short (minutes). The buildbot activity and logs can be browsed online.

Regression tests

Yade contains two types of regression tests, some are unit tests while others are testing more complex simulations. Altough both types can be considered regression tests, the usage is that we name the first simply “regression tests”, while the latest are called “check tests”. Both series of tests can be ran at yade startup by passing the options “test” or “check”

yade --test
yade --check

Unit regression tests

Unit regression tests are testing the output of individual functors and engines in well defined conditions. They are defined in the folder py/tests/. The purpose of unit testing is to make sure that the behaviour of the most important classes remains correct during code development. Since they test classes one by one, unit tests can’t detect problems coming from the interaction between different engines in a typical simulation. That is why check tests have been introduced.

Check tests

Check tests perform comparisons of simulation results between different versions of yade, as discussed here. They differ with regression tests in the sense that they simulate more complex situations and combinations of different engines, and usually don’t have a mathematical proof (though there is no restriction on the latest). They compare the values obtained in version N with values obtained in a previous version or any other “expected” results. The reference values must be hardcoded in the script itself or in data files provided with the script. Check tests are based on regular yade scripts, so that users can easily commit their own scripts to trunk in order to get some automatized testing after commits from other developers.

Since the check tests history will be mostly based on standard output generated by “yade —check”, a meaningfull checkTest should include some “print” command telling if something went wrong. If the script itself fails for some reason and can’t generate an output, the log will contain “scriptName failure”. If the script defines differences on obtained and awaited data, it should print some useful information about the problem and increase the value of global variable resultStatus. After this occurs, the automatic test will stop the execution with error message.

An example check test can be found in It shows results comparison, output, and how to define the path to data files using “checksPath”. Users are encouraged to add their own scripts into the scripts/test/checks/ folder. Discussion of some specific checktests design in users question is welcome. Note that re-compiling is required before that added scripts can be launched by “yade —check” (or direct changes have to be performed in “lib” subfolders). A check test should never need more than a few seconds to run. If your typical script needs more, try and reduce the number of element or the number of steps.


The following rules that should be respected; documentation is treated separately.

  • general
    • C++ source files have .hpp and .cpp extensions (for headers and implementation, respectively).
    • All header files should have the #pragma once multiple-inclusion guard.
    • Try to avoid using namespace in header files.
    • Use tabs for indentation. While this is merely visual in c++, it has semantic meaning in python; inadverently mixing tabs and spaces can result in syntax errors.
  • capitalization style
    • Types should be always capitalized. Use CamelCase for composed names (GlobalEngine). Underscores should be used only in special cases, such as functor names.
    • Class data members and methods must not be capitalized, composed names should use use lowercased camelCase (glutSlices). The same applies for functions in python modules.
    • Preprocessor macros are uppercase, separated by underscores; those that are used outside the core take (with exceptions) the form YADE_*, such as YADE_CLASS_BASE_DOC_* macro family.
  • programming style
    • Be defensive, if it has no significant performance impact. Use assertions abundantly: they don’t affect performance (in the optimized build) and make spotting error conditions much easier.
    • Use YADE_CAST and YADE_PTR_CAST where you want type-check during debug builds, but fast casting in optimized build.
    • Initialize all class variables in the default constructor. This avoids bugs that may manifest randomly and are difficult to fix. Initializing with NaN’s will help you find otherwise unitialized variable. (This is taken care of by YADE_CLASS_BASE_DOC_* macro family macros for user classes)

Class naming

Although for historial reasons the naming scheme is not completely consistent, these rules should be obeyed especially when adding a new class.

GlobalEngines and PartialEngines
GlobalEngines should be named in a way suggesting that it is a performer of certain action (like ForceResetter, InsertionSortCollider, Recorder); if this is not appropriate, append the Engine to the characteristics (GravityEngine). PartialEngines have no special naming convention different from GlobalEngines.
Names of all dispatchers end in Dispatcher. The name is composed of type it creates or, in case it doesn’t create any objects, its main characteristics. Currently, the following dispatchers [1] are defined:
dispatcher arity dispatch types created type functor type functor prefix
BoundDispatcher 1 Shape Bound BoundFunctor Bo1
IGeomDispatcher 2 (symetric) 2 × Shape IGeom IGeomFunctor Ig2
IPhysDispatcher 2 (symetric) 2 × Material IPhys IPhysFunctor Ip2
LawDispatcher 2 (asymetric) IGeom IPhys (none) LawFunctor Law2

Respective abstract functors for each dispatchers are BoundFunctor, IGeomFunctor, IPhysFunctor and LawFunctor.


Functor name is composed of 3 parts, separated by underscore.

  1. prefix, composed of abbreviated functor type and arity (see table above)
  2. Types entering the dispatcher logic (1 for unary and 2 for binary functors)
  3. Return type for functors that create instances, simple characteristics for functors that don’t create instances.

To give a few examples:

[1]Not considering OpenGL dispatchers, which might be replaced by regular virtual functions in the future.


Documenting code properly is one of the most important aspects of sustained development.

Read it again.

Most code in research software like Yade is not only used, but also read, by developers or even by regular users. Therefore, when adding new class, always mention the following in the documentation:

  • purpose
  • details of the functionality, unless obvious (algorithms, internal logic)
  • limitations (by design, by implementation), bugs
  • bibliographical reference, if using non-trivial published algorithms (see below)
  • references to other related classes
  • hyperlinks to bugs, blueprints, wiki or mailing list about this particular feature.

As much as it is meaningful, you should also

  • update any other documentation affected
  • provide a simple python script demonstrating the new functionality in scripts/test.

Sphinx documentation

Most c++ classes are wrapped in Python, which provides good introspection and interactive documentation (try writing Material? in the ipython prompt; or help(CpmState)).

Syntax of documentation is ReST (reStructuredText, see reStructuredText Primer). It is the same for c++ and python code.

  • Documentation of c++ classes exposed to python is given as 3rd argument to YADE_CLASS_BASE_DOC_* macro family introduced below.
  • Python classes/functions are documented using regular python docstrings. Besides explaining functionality, meaning and types of all arguments should also be documented. Short pieces of code might be very helpful. See the utils module for an example.

In addition to standard ReST syntax, yade provides several shorthand macros:


creates hyperlink to referenced term, for instance:


becomes CpmMat; link name and target can be different:

:yref:`Material used in the CPM model<CpmMat>`

yielding Material used in the CPM model.


creates hyperlink to file within the source tree (to its latest version in the repository), for instance core/Cell.hpp. Just like with :yref:, alternate text can be used with

:ysrc:`Link text<target/file>`

like this.

is used in attribute description for those that should not be provided by the user, but are auto-computed instead; |ycomp| expands to (auto-computed).
marks attributes that are periodically updated, being subset of the previous. |yupdate| expands to (auto-updated).

delimits inline math expressions; they will be replaced by:


and rendered via LaTeX. To write a single dollar sign, escape it with backslash \$.

Displayed mathematics (standalone equations) can be inserted as explained in Math support in Sphinx.

Bibliographical references

As in any scientific documentation, references to publications are very important. To cite an article, add it to BibTeX file in doc/references.bib, using the BibTeX format. Please adhere to the following conventions:

  1. Keep entries in the form Author2008 (Author is the first author), Author2008b etc if multiple articles from one author;
  2. Try to fill mandatory fields for given type of citation;
  3. Do not use \'{i} funny escapes for accents, since they will not work with the HTML output; put everything in straight utf-8.

In your docstring, the Author2008 article can be cited by [Author2008]_; for example:

According to [Allen1989]_, the integration scheme …

will be rendered as

According to [Allen1989], the integration scheme …

Separate class/function documentation

Some c++ might have long or content-rich documentation, which is rather inconvenient to type in the c++ source itself as string literals. Yade provides a way to write documentation separately in py/ file: it is executed after loading c++ plugins and can set __doc__ attribute of any object directly, overwriting docstring from c++. In such (exceptional) cases:

  1. Provide at least a brief description of the class in the c++ code nevertheless, for people only reading the code.

  2. Add notice saying “This class is documented in detail in the py/ file”.

  3. Add documentation to py/ in this way:

            This is the docstring for YourClass.
            Class, methods and functions can be documented this way.
            .. note:: It can use any syntax features you like.


Boost::python embeds function signatures in the docstring (before the one provided by the user). Therefore, before creating separate documentation of your function, have a look at its __doc__ attribute and copy the first line (and the blank line afterwards) in the separate docstring. The first line is then used to create the function signature (arguments and return value).

Internal c++ documentation

doxygen was used for automatic generation of c++ code. Since user-visible classes are defined with sphinx now, it is not meaningful to use doxygen to generate overall documentation. However, take care to document well internal parts of code using regular comments, including public and private data members.

Support framework

Besides the framework provided by the c++ standard library (including STL), boost and other dependencies, yade provides its own specific services.


Shared pointers

Yade makes extensive use of shared pointers shared_ptr. [2] Although it probably has some performance impacts, it greatly simplifies memory management, ownership management of c++ objects in python and so forth. To obtain raw pointer from a shared_ptr, use its get() method; raw pointers should be used in case the object will be used only for short time (during a function call, for instance) and not stored anywhere.

[2]Either boost::shared_ptr or tr1::shared_ptr is used, but it is always imported with the using statement so that unqualified shared_ptr can be used.

Python defines thin wrappers for most c++ Yade classes (for all those registered with YADE_CLASS_BASE_DOC_* macro family and several others), which can be constructed from shared_ptr; in this way, Python reference counting blends with the shared_ptr reference counting model, preventing crashes due to python objects pointing to c++ objects that were destructed in the meantime.


Frequently, pointers have to be typecast; there is choice between static and dynamic casting.

  • dynamic_cast (dynamic_pointer_cast for a shared_ptr) assures cast admissibility by checking runtime type of its argument and returns NULL if the cast is invalid; such check obviously costs time. Invalid cast is easily caught by checking whether the pointer is NULL or not; even if such check (e.g. assert) is absent, dereferencing NULL pointer is easily spotted from the stacktrace (debugger output) after crash. Moreover, shared_ptr checks that the pointer is non-NULL before dereferencing in debug build and aborts with “Assertion `px!=0’ failed.” if the check fails.
  • static_cast is fast but potentially dangerous (static_pointer_cast for shared_ptr). Static cast will return non-NULL pointer even if types don’t allow the cast (such as casting from State* to Material*); the consequence of such cast is interpreting garbage data as instance of the class cast to, leading very likely to invalid memory access (segmentation fault, “crash” for short).

To have both speed and safety, Yade provides 2 macros:

expands to static_cast in optimized builds and to dynamic_cast in debug builds.
expands to static_pointer_cast in optimized builds and to dynamic_pointer_cast in debug builds.

Basic numerics

The floating point type to use in Yade Real, which is by default typedef for double. [3]

Yade uses the Eigen library for computations. It provides classes for 2d and 3d vectors, quaternions and 3x3 matrices templated by number type; their specialization for the Real type are typedef’ed with the “r” suffix, and occasionally useful integer types with the “i” suffix:

  • Vector2r, Vector2i
  • Vector3r, Vector3i
  • Quaternionr
  • Matrix3r

Yade additionally defines a class named Se3r, which contains spatial position (Vector3r Se3r::position) and orientation (Quaternionr Se3r::orientation), since they are frequently used one with another, and it is convenient to pass them as single parameter to functions.

[3]Historically, it was thought that Yade could be also run with single precision based on build-time parameter; it turned out however that the impact on numerical stability was such disastrous that this option is not available now. There is, however, QUAD_PRECISION parameter to scons, which will make Real a typedef for long double (extended precision; quad precision in the proper sense on IA64 processors); this option is experimental and is unlikely to be used in near future, though.

Eigen provides full rich linear algebra functionality. Some code firther uses the [cgal] library for computational geometry.

In Python, basic numeric types are wrapped and imported from the minieigen module; the types drop the r type qualifier at the end, the syntax is otherwise similar. Se3r is not wrapped at all, only converted automatically, rarely as it is needed, from/to a (Vector3,Quaternion) tuple/list.

# cross product
Yade [1]: Vector3(1,2,3).cross(Vector3(0,0,1))
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 Vector3(1,2,3).cross(Vector3(0,0,1))

NameError: name 'Vector3' is not defined

# construct quaternion from axis and angle
Yade [2]: Quaternion(Vector3(0,0,1),pi/2)
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 Quaternion(Vector3(0,0,1),pi/2)

NameError: name 'Quaternion' is not defined


Quaternions are internally stored as 4 numbers. Their usual human-readable representation is, however, (normalized) axis and angle of rotation around that axis, and it is also how they are input/output in Python. Raw internal values can be accessed using the [0][3] element access (or .W(), .X(), .Y() and .Z() methods), in both c++ and Python.

Run-time type identification (RTTI)

Since serialization and dispatchers need extended type and inheritance information, which is not sufficiently provided by standard RTTI. Each yade class is therefore derived from Factorable and it must use macro to override its virtual functions providing this extended RTTI:

YADE_CLASS_BASE_DOC(Foo,Bar Baz,"Docstring) creates the following virtual methods (mediated via the REGISTER_CLASS_AND_BASE macro, which is not user-visible and should not be used directly):

  • std::string getClassName() returning class name (Foo) as string. (There is the typeid(instanceOrType).name() standard c++ construct, but the name returned is compiler-dependent.)
  • unsigned getBaseClassNumber() returning number of base classes (in this case, 2).
  • std::string getBaseClassName(unsigned i=0) returning name of i-th base class (here, Bar for i=0 and Baz for i=1).


RTTI relies on virtual functions; in order for virtual functions to work, at least one virtual method must be present in the implementation (.cpp) file. Otherwise, virtual method table (vtable) will not be generated for this class by the compiler, preventing virtual methods from functioning properly.

Some RTTI information can be accessed from python:

Yade [3]: yade.system.childClasses('Shape')

Yade [4]: Sphere().name            ## getClassName()
AttributeError                            Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 Sphere().name            ## getClassName()

AttributeError: 'Sphere' object has no attribute 'name'


Serialization serves to save simulation to file and restore it later. This process has several necessary conditions:

  • classes know which attributes (data members) they have and what are their names (as strings);
  • creating class instances based solely on its name;
  • knowing what classes are defined inside a particular shared library (plugin).

This functionality is provided by 3 macros and 4 optional methods; details are provided below.

Serializable::preLoad, Serializable::preSave, Serializable::postLoad, Serializable::postSave

Prepare attributes before serialization (saving) or deserialization (loading) or process them after serialization or deserialization.

See Attribute registration.


Inside the class declaration (i.e. in the .hpp file within the class Foo { /* */}; block). See Attribute registration.

Enumerate class attributes that should be saved and loaded; associate each attribute with its literal name, which can be used to retrieve it. See YADE_CLASS_BASE_DOC_* macro family.

Additionally documents the class in python, adds methods for attribute access from python, and documents each attribute.


In header file, but after the class declaration block. See Class factory.

Associate literal name of the class with functions that will create its new instance (ClassFactory).


In the implementation .cpp file. See Plugin registration.

Declare what classes are declared inside a particular plugin at time the plugin is being loaded (yade startup).

Attribute registration

All (serializable) types in Yade are one of the following:

  • Type deriving from Serializable, which provide information on how to serialize themselves via overriding the Serializable::registerAttributes method; it declares data members that should be serialzed along with their literal names, by which they are identified. This method then invokes registerAttributes of its base class (until Serializable itself is reached); in this way, derived classes properly serialize data of their base classes.

    This funcionality is hidden behind the macro YADE_CLASS_BASE_DOC_* macro family used in class declaration body (header file), which takes base class and list of attributes:

    YADE_CLASS_BASE_DOC_ATTRS(ThisClass,BaseClass,"class documentation",((type1,attribute1,initValue1,,"Documentation for attribute 1"))((type2,attribute2,initValue2,,"Documentation for attribute 2"));

    Note that attributes are encodes in double parentheses, not separated by commas. Empty attribute list can be given simply by YADE_CLASS_BASE_DOC_ATTRS(ThisClass,BaseClass,"documentation",) (the last comma is mandatory), or by omiting ATTRS from macro name and last parameter altogether.

  • Fundamental type: strings, various number types, booleans, Vector3r and others. Their “handlers” (serializers and deserializers) are defined in lib/serialization.

  • Standard container of any serializable objects.

  • Shared pointer to serializable object.

Yade uses the excellent boost::serialization library internally for serialization of data.


YADE_CLASS_BASE_DOC_ATTRS also generates code for attribute access from python; this will be discussed later. Since this macro serves both purposes, the consequence is that attributes that are serialized can always be accessed from python.

Yade also provides callback for before/after (de) serialization, virtual functions Serializable::preProcessAttributes and Serializable::postProcessAttributes, which receive one bool deserializing argument (true when deserializing, false when serializing). Their default implementation in Serializable doesn’t do anything, but their typical use is:

  • converting some non-serializable internal data structure of the class (such as multi-dimensional array, hash table, array of pointers) into a serializable one (pre-processing) and fill this non-serializable structure back after deserialization (post-processing); for instance, InteractionContainer uses these hooks to ask its concrete implementation to store its contents to a unified storage (vector<shared_ptr<Interaction> >) before serialization and to restore from it after deserialization.
  • precomputing non-serialized attributes from the serialized values; e.g. Facet computes its (local) edge normals and edge lengths from vertices’ coordinates.

Class factory

Each serializable class must use REGISTER_SERIALIZABLE, which defines function to create that class by ClassFactory. ClassFactory is able to instantiate a class given its name (as string), which is necessary for deserialization.

Although mostly used internally by the serialization framework, programmer can ask for a class instantiation using shared_ptr<Factorable> f=ClassFactory::instance().createShared("ClassName");, casting the returned shared_ptr<Factorable> to desired type afterwards. Serializable itself derives from Factorable, i.e. all serializable types are also factorable (It is possible that different mechanism will be in place if boost::serialization is used, though.)

Plugin registration

Yade loads dynamic libraries containing all its functionality at startup. ClassFactory must be taught about classes each particular file provides. YADE_PLUGIN serves this purpose and, contrary to YADE_CLASS_BASE_DOC_* macro family, must be place in the implementation (.cpp) file. It simple enumerates classes that are provided by this file:



You must use parentheses around the class name even if there is only one (preprocessor limitation): YADE_PLUGIN((classFoo));. If there is no class in this file, do not use this macro at all.

Internally, this macro creates function registerThisPluginClasses_ declared specially as __attribute__((constructor)) (see GCC Function Attributes); this attributes makes the function being executed when the plugin is loaded via dlopen from ClassFactory::load(...). It registers all factorable classes from that file in the Class factory.


Classes that do not derive from Factorable, such as Shop or SpherePack, are not declared with YADE_PLUGIN.

This is an example of a serializable class header:

/*! Homogeneous gravity field; applies gravity×mass force on all bodies. */
class GravityEngine: public GlobalEngine{
                virtual void action();
        // registering class and its base for the RTTI system
                // documentation visible from python and generated reference documentation
                "Homogeneous gravity field; applies gravity×mass force on all bodies.",
                // enumerating attributes here, include documentation
                ((Vector3r,gravity,Vector3r::ZERO,"acceleration, zero by default [kgms⁻²]"))
// registration function for ClassFactory

and this is the implementation:


// registering the plugin

void GravityEngine::action(){
        /* do the work here */

We can create a mini-simulation (with only one GravityEngine):

Yade [5]: O.engines=[GravityEngine(gravity=Vector3(0,0,-9.81))]
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 O.engines=[GravityEngine(gravity=Vector3(0,0,-9.81))]

NameError: name 'Vector3' is not defined

Yade [6]:'abc.xml')

and the XML looks like this:

<?xml version="1.0" encoding="UTF-8" standalone="yes" ?>
<!DOCTYPE boost_serialization>
<boost_serialization signature="serialization::archive" version="10">
<scene class_id="0" tracking_level="0" version="1">
	<px class_id="1" tracking_level="1" version="0" object_id="_0">
		<Serializable class_id="2" tracking_level="1" version="0" object_id="_1"></Serializable>
		<tags class_id="3" tracking_level="0" version="0">
		<engines class_id="4" tracking_level="0" version="0">
		<bodies class_id="5" tracking_level="0" version="1">
			<px class_id="6" tracking_level="1" version="0" object_id="_2">
				<Serializable object_id="_3"></Serializable>
				<body class_id="7" tracking_level="0" version="0">
		<interactions class_id="8" tracking_level="0" version="1">
			<px class_id="9" tracking_level="1" version="0" object_id="_4">
				<Serializable object_id="_5"></Serializable>
				<interaction class_id="10" tracking_level="0" version="0">
		<energy class_id="11" tracking_level="0" version="1">
			<px class_id="12" tracking_level="1" version="0" object_id="_6">
				<Serializable object_id="_7"></Serializable>
				<energies class_id="13" tracking_level="0" version="0">
				<names class_id="14" tracking_level="0" version="0">
		<materials class_id="16" tracking_level="0" version="0">
		<bound class_id="17" tracking_level="0" version="1">
			<px class_id="-1"></px>
		<cell class_id="19" tracking_level="0" version="1">
			<px class_id="20" tracking_level="1" version="0" object_id="_8">
				<Serializable object_id="_9"></Serializable>
				<trsf class_id="21" tracking_level="0" version="0">
		<miscParams class_id="22" tracking_level="0" version="0">
		<dispParams class_id="23" tracking_level="0" version="0">


Since XML files closely reflect implementation details of Yade, they will not be compatible between different versions. Use them only for short-term saving of scenes. Python is the high-level description Yade uses.

Python attribute access

The macro YADE_CLASS_BASE_DOC_* macro family introduced above is (behind the scenes) also used to create functions for accessing attributes from Python. As already noted, set of serialized attributes and set of attributes accessible from Python are identical. Besides attribute access, these wrapper classes imitate also some functionality of regular python dictionaries:

Yade [7]: s=Sphere()

Yade [8]: s.radius              ## read-access
Out[8]: nan

Yade [9]: s.radius=4.           ## write access

Yade [10]: s.dict().keys()              ## show all available keys
Out[10]: ['color', 'highlight', 'wire', 'radius']

Yade [11]: for k in s.dict().keys(): print s.dict()[k]  ## iterate over keys, print their values

Yade [12]: s.dict()['radius']             ## same as: 'radius' in s.keys()
Out[12]: 4.0

Yade [13]: s.dict()                       ## show dictionary of both attributes and values
Out[13]: {'color': Vector3(1,1,1), 'highlight': False, 'radius': 4.0, 'wire': False}

YADE_CLASS_BASE_DOC_* macro family

There is several macros that hide behind them the functionality of Sphinx documentation, Run-time type identification (RTTI), Attribute registration, Python attribute access, plus automatic attribute initialization and documentation. They are all defined as shorthands for base macro YADE_CLASS_BASE_DOC_ATTRS_INIT_CTOR_PY with some arguments left out. They must be placed in class declaration’s body (.hpp file):

#define YADE_CLASS_BASE_DOC(klass,base,doc) \
#define YADE_CLASS_BASE_DOC_ATTRS(klass,base,doc,attrs) \
#define YADE_CLASS_BASE_DOC_ATTRS_CTOR(klass,base,doc,attrs,ctor) \
#define YADE_CLASS_BASE_DOC_ATTRS_CTOR_PY(klass,base,doc,attrs,ctor,py) \
#define YADE_CLASS_BASE_DOC_ATTRS_INIT_CTOR_PY(klass,base,doc,attrs,init,ctor,py) \

Expected parameters are indicated by macro name components separated with underscores. Their meaning is as follows:

(unquoted) name of this class (used for RTTI and python)
(unquoted) name of the base class (used for RTTI and python)

docstring of this class, written in the ReST syntax. This docstring will appear in generated documentation (such as CpmMat). It can be as long as necessary, but sequences interpreted by c++ compiler must be properly escaped (therefore some backslashes must be doubled, like in \sigma=\epsilon E:

":math:`\\sigma=\\epsilon E"

Use \n and \t for indentation inside the docstring. Hyperlink the documentation abundantly with yref (all references to other classes should be hyperlinks).

See Sphinx documentation for syntax details.


Attribute must be written in the form of parethesized list:

((type1,attr1,initValue1,attrFlags,"Attribute 1 documentation"))
((type2,attr2,,,"Attribute 2 documentation"))  // initValue and attrFlags unspecified

This will expand to

  1. data members declaration in c++ (note that all attributes are public):

    public: type1 attr1;
            type2 attr2;
  2. Initializers of the default (argument-less) constructor, for attributes that have non-empty initValue:

    Klass(): attr1(initValue1), attr2() { /* constructor body */ }

    No initial value will be assigned for attribute of which initial value is left empty (as is for attr2 in the above example). Note that you still have to write the commas.

  3. Registration of the attribute in the serialization system (unless disabled by attrFlags – see below)

  4. Registration of the attribute in python (unless disabled by attrFlags), so that it can be accessed as klass().name1.

    The attribute is read-write by default, see attrFlags to change that.

    This attribute will carry the docstring provided, along with knowledge of the initial value. You can add text description to the default value using the comma operator of c++ and casting the char* to (void):

    ((Real,dmgTau,((void)"deactivated if negative",-1),,"Characteristic time for normal viscosity. [s]"))

    leading to CpmMat::dmgTau.

    The attribute is registered via boost::python::add_property specifying return_by_value policy rather than return_internal_reference, which is the default when using def_readwrite. The reason is that we need to honor custom converters for those values; see note in Custom converters for details.

Attribute flags

By default, an attribute will be serialized and will be read-write from python. There is a number of flags that can be passed as the 4th argument (empty by default) to change that:

  • Attr::noSave avoids serialization of the attribute (while still keeping its accessibility from Python)
  • Attr::readonly makes the attribute read-only from Python
  • Attr::triggerPostLoad will trigger call to postLoad function to handle attribute change after its value is set from Python; this is to ensure consistency of other precomputed data which depend on this value (such as Cell.trsf and such)
  • Attr::hidden will not expose the attribute to Python at all
  • Attr::noResize will not permit changing size of the array from Python [not yet used]

Flags can be combined as usual using bitwise disjunction | (such as Attr::noSave | Attr::readonly), though in such case the value should be parenthesized to avoid a warning with some compilers (g++ specifically), i.e. (Attr::noSave | Attr::readonly).

Currently, the flags logic handled at runtime; that means that even for attributes with Attr::noSave, their serialization template must be defined (although it will never be used). In the future, the implementation might be template-based, avoiding this necessity.


List of deprecated attribute names. The syntax is

((oldName1,newName1,"Explanation why renamed etc."))
((oldName2,newName2,"! Explanation why removed and what to do instaed."))

This will make accessing oldName1 attribute from Python return value of newName, but displaying warning message about the attribute name change, displaying provided explanation. This happens whether the access is read or write.

If the explanation’s first character is ! (bang), the message will be displayed upon attribute access, but exception will be thrown immediately. Use this in cases where attribute is no longer meaningful or was not straightforwardsly replaced by another, but more complex adaptation of user’s script is needed. You still have to give newName2, although its value will never be used – you can use any variable you like, but something must be given for syntax reasons).


Due to compiler limitations, this feature only works if Yade is compiled with gcc >= 4.4. In the contrary case, deprecated attribute functionality is disabled, even if such attributes are declared.


Parethesized list of the form:

((attr3,value3)) ((attr4,value4))

which will be expanded to initializers in the default ctor:

Klass(): /* attributes declared with the attrs argument */ attr4(value4), attr5(value5) { /* constructor body */ }

The purpose of this argument is to make it possible to initialize constants and references (which are not declared as attributes using this macro themselves, but separately), as that cannot be done in constructor body. This argument is rarely used, though.


will be put directly into the generated constructor’s body. Mostly used for calling createIndex(); in the constructor.


The code must not contain commas ouside parentheses (since preprocessor uses commas to separate macro arguments). If you need complex things at construction time, create a separate init() function and call it from the constructor instead.


will be appeneded directly after generated python code that registers the class and all its attributes. You can use it to access class methods from python, for instance, to override an existing attribute with the same name etc:

.def_readonly("omega",&CpmPhys::omega,"Damage internal variable")
.def_readonly("Fn",&CpmPhys::Fn,"Magnitude of normal force.")

def_readonly will not work for custom types (such as std::vector), as it bypasses conversion registry; see Custom converters for details.

Special python constructors

The Python wrapper automatically create constructor that takes keyword (named) arguments corresponding to instance attributes; those attributes are set to values provided in the constructor. In some cases, more flexibility is desired (such as InteractionLoop, which takes 3 lists of functors). For such cases, you can override the function Serializable::pyHandleCustomCtorArgs, which can arbitrarily modify the new (already existing) instance. It should modify in-place arguments given to it, as they will be passed further down to the routine which sets attribute values. In such cases, you should document the constructor:

.. admonition:: Special constructor

        Constructs from lists of …

which then appears in the documentation similar to InteractionLoop.

Static attributes

Some classes (such as OpenGL functors) are instantiated automatically; since we want their attributes to be persistent throughout the session, they are static. To expose class with static attributes, use the YADE_CLASS_BASE_DOC_STATICATTRS macro. Attribute syntax is the same as for YADE_CLASS_BASE_DOC_ATTRS:

class SomeClass: public BaseClass{
        YADE_CLASS_BASE_DOC_STATICATTRS(SomeClass,BaseClass,"Documentation of SomeClass",
                ((Type1,attr1,default1,"doc for attr1"))
                ((Type2,attr2,default2,"doc for attr2"))

additionally, you have to allocate memory for static data members in the .cpp file (otherwise, error about undefined symbol will appear when the plugin is loaded):

There is no way to expose class that has both static and non-static attributes using YADE_CLASS_BASE_* macros. You have to expose non-static attributes normally and wrap static attributes separately in the py parameter.

Returning attribute by value or by reference

When attribute is passed from c++ to python, it can be passed either as

  • value: new python object representing the original c++ object is constructed, but not bound to it; changing the python object doesn’t modify the c++ object, unless explicitly assigned back to it, where inverse conversion takes place and the c++ object is replaced.
  • reference: only reference to the underlying c++ object is given back to python; modifying python object will make the c++ object modified automatically.

The way of passing attributes given to YADE_CLASS_BASE_DOC_ATTRS in the attrs parameter is determined automatically in the following manner:

  • Vector3, Vector3i, Vector2, Vector2i, Matrix3 and Quaternion objects are passed by reference. For instance::


    will assign correct value to x component of position, without changing the other ones.

  • Yade classes (all that use shared_ptr when declared in python: all classes deriving from Serializable declared with YADE_CLASS_BASE_DOC_*, and some others) are passed as references (technically speaking, they are passed by value of the shared_ptr, but by virtue of its sharedness, they appear as references). For instance::


    will change damping parameter on the original engine object, not on its copy.

  • All other types are passed by value. This includes, most importantly, sequence types declared in Custom converters, such as std::vector<shared_ptr<Engine> >. For this reason, ::


    will not work as expected; it will replace 5th element of a copy of the sequence, and this change will not propagate back to c++.

Multiple dispatch

Multiple dispatch is generalization of virtual methods: a Dispatcher decides based on type(s) of its argument(s) which of its Functors to call. Numer of arguments (currently 1 or 2) determines arity of the dispatcher (and of the functor): unary or binary. For example:


creates InsertionSortCollider, which internally contains Collider.boundDispatcher, a BoundDispatcher (a Dispatcher), with 2 functors; they receive Sphere or Facet instances and create Aabb. This code would look like this in c++:

shared_ptr<InsertionSortCollider> collider=(new InsertionSortCollider);
collider->boundDispatcher->add(new Bo1_Sphere_Aabb());
collider->boundDispatcher->add(new Bo1_Facet_Aabb());

There are currenly 4 predefined dispatchers (see dispatcher-names) and corresponding functor types. They are inherit from template instantiations of Dispatcher1D or Dispatcher2D (for functors, Functor1D or Functor2D). These templates themselves derive from DynlibDispatcher (for dispatchers) and FunctorWrapper (for functors).

Example: IGeomDispatcher

Let’s take (the most complicated perhaps) IGeomDispatcher. IGeomFunctor, which is dispatched based on types of 2 Shape instances (a Functor), takes a number of arguments and returns bool. The functor “call” is always provided by its overridden Functor::go method; it always receives the dispatched instances as first argument(s) (2 × const shared_ptr<Shape>&) and a number of other arguments it needs:

class IGeomFunctor: public Functor2D<
   bool,                                 //return type
   TYPELIST_7(const shared_ptr<Shape>&,  // 1st class for dispatch
      const shared_ptr<Shape>&,          // 2nd class for dispatch
      const State&,                      // other arguments passed to ::go
      const State&,                      // …
      const Vector3r&,                   // …
      const bool&,                       // …
      const shared_ptr<Interaction>&     // …

The dispatcher is declared as follows:

class IGeomDispatcher: public Dispatcher2D<
   Shape,                       // 1st class for dispatch
   Shape,                       // 2nd class for dispatch
   IGeomFunctor,  // functor type
   bool,                        // return type of the functor

   // follow argument types for functor call
   // they must be exactly the same as types
   // given to the IGeomFunctor above.
   TYPELIST_7(const shared_ptr<Shape>&,
      const shared_ptr<Shape>&,
      const State&,
      const State&,
      const Vector3r&,
      const bool &,
      const shared_ptr<Interaction>&

   // handle symetry automatically
   // (if the dispatcher receives Sphere+Facet,
   // the dispatcher might call functor for Facet+Sphere,
   // reversing the arguments)
{ /* … */ }

Functor derived from IGeomFunctor must then

  • override the ::go method with appropriate arguments (they must match exactly types given to TYPELIST_* macro);
  • declare what types they should be dispatched for, and in what order if they are not the same.
class Ig2_Facet_Sphere_ScGeom: public IGeomFunctor{

   // override the IGeomFunctor::go
   //   (it is really inherited from FunctorWrapper template,
   //    therefore not declare explicitly in the
   //    IGeomFunctor declaration as such)
   // since dispatcher dispatches only for declared types
   //   (or types derived from them), we can do
   //   static_cast<Facet>(shape1) and static_cast<Sphere>(shape2)
   //   in the ::go body, without worrying about types being wrong.
   virtual bool go(
      // objects for dispatch
      const shared_ptr<Shape>& shape1, const shared_ptr<Shape>& shape2,
      // other arguments
      const State& state1, const State& state2, const Vector3r& shift2,
      const bool& force, const shared_ptr<Interaction>& c
   /* … */

   // this declares the type we want to be dispatched for, matching
   //   first 2 arguments to ::go and first 2 classes in TYPELIST_7 above
   //   shape1 is a Facet and shape2 is a Sphere
   //   (or vice versa, see lines below)

   // declare how to swap the arguments
   //   so that we can receive those as well
   /* … */

Dispatch resolution

The dispatcher doesn’t always have functors that exactly match the actual types it receives. In the same way as virtual methods, it tries to find the closest match in such way that:

  1. the actual instances are derived types of those the functor accepts, or exactly the accepted types;
  2. sum of distances from actual to accepted types is sharp-minimized (each step up in the class hierarchy counts as 1)

If no functor is able to accept given types (first condition violated) or multiple functors have the same distance (in condition 2), an exception is thrown.

This resolution mechanism makes it possible, for instance, to have a hierarchy of ScGeom classes (for different combination of shapes), but only provide a LawFunctor accepting ScGeom, rather than having different laws for each shape combination.


Performance implications of dispatch resolution are relatively low. The dispatcher lookup is only done once, and uses fast lookup matrix (1D or 2D); then, the functor found for this type(s) is cached within the Interaction (or Body) instance. Thus, regular functor call costs the same as dereferencing pointer and calling virtual method. There is blueprint to avoid virtual function call as well.


At the beginning, the dispatch matrix contains just entries exactly matching given functors. Only when necessary (by passing other types), appropriate entries are filled in as well.

Indexing dispatch types

Classes entering the dispatch mechanism must provide for fast identification of themselves and of their parent class. [4] This is called class indexing and all such classes derive from Indexable. There are top-level Indexables (types that the dispatchers accept) and each derived class registers its index related to this top-level Indexable. Currently, there are:

Top-level Indexable used by
Shape BoundFunctor, IGeomDispatcher
Material IPhysDispatcher
IPhys LawDispatcher
IGeom LawDispatcher

The top-level Indexable must use the REGISTER_INDEX_COUNTER macro, which sets up the machinery for identifying types of derived classes; they must then use the REGISTER_CLASS_INDEX macro and call createIndex() in their constructor. For instance, taking the Shape class (which is a top-level Indexable):

// derive from Indexable
class Shape: public Serializable, public Indexable {
   // never call createIndex() in the top-level Indexable ctor!
   /* … */

   // allow index registration for classes deriving from ``Shape``

Now, all derived classes (such as Sphere or Facet) use this:

class Sphere: public Shape{
   /* … */
      /* … */,
      // this is the CTOR argument
         // important; assigns index to the class at runtime
   // register index for this class, and give name of the immediate parent class
   //    (i.e. if there were a class deriving from Sphere, it would use
   //     REGISTER_CLASS_INDEX(SpecialSphere,Sphere),
   //     not REGISTER_CLASS_INDEX(SpecialSphere,Shape)!)

At runtime, each class within the top-level Indexable hierarchy has its own unique numerical index. These indices serve to build the dispatch matrix for each dispatcher.

[4]The functionality described in Run-time type identification (RTTI) serves a different purpose (serialization) and would hurt the performance here. For this reason, classes provide numbers (indices) in addition to strings.

Inspecting dispatch in python

If there is a need to debug/study multiple dispatch, python provides convenient interface for this low-level functionality.

We can inspect indices with the dispIndex property (note that the top-level indexable Shape has negative (invalid) class index; we purposively didn’t call createIndex in its constructor):

Yade [14]: Sphere().dispIndex, Facet().dispIndex, Wall().dispIndex
Out[14]: (1, 7, 16)

Yade [15]: Shape().dispIndex                  # top-level indexable
Out[15]: -1

Dispatch hierarchy for a particular class can be shown with the dispHierarchy() function, returning list of class names: 0th element is the instance itself, last element is the top-level indexable (again, with invalid index); for instance:

Yade [16]: ScGeom().dispHierarchy()       # parent class of all other ScGeom_ classes
Out[16]: ['ScGeom', 'GenericSpheresContact', 'IGeom']

Yade [17]: ScGridCoGeom().dispHierarchy(), ScGeom6D().dispHierarchy(), CylScGeom().dispHierarchy()
(['ScGridCoGeom', 'ScGeom6D', 'ScGeom', 'GenericSpheresContact', 'IGeom'],
 ['ScGeom6D', 'ScGeom', 'GenericSpheresContact', 'IGeom'],
 ['CylScGeom', 'ScGeom', 'GenericSpheresContact', 'IGeom'])

Yade [18]: CylScGeom().dispHierarchy(names=False)   # show numeric indices instead
Out[18]: [4, 1, 0, -1]

Dispatchers can also be inspected, using the .dispMatrix() method:

Yade [19]: ig=IGeomDispatcher([
   ....:    Ig2_Sphere_Sphere_ScGeom(),
   ....:    Ig2_Facet_Sphere_ScGeom(),
   ....:    Ig2_Wall_Sphere_ScGeom()
   ....: ])

Yade [20]: ig.dispMatrix()
{('Facet', 'Sphere'): 'Ig2_Facet_Sphere_ScGeom',
 ('Sphere', 'Facet'): 'Ig2_Facet_Sphere_ScGeom',
 ('Sphere', 'Sphere'): 'Ig2_Sphere_Sphere_ScGeom',
 ('Sphere', 'Wall'): 'Ig2_Wall_Sphere_ScGeom',
 ('Wall', 'Sphere'): 'Ig2_Wall_Sphere_ScGeom'}

Yade [21]: ig.dispMatrix(False)          # don't convert to class names
{(1, 1): 'Ig2_Sphere_Sphere_ScGeom',
 (1, 7): 'Ig2_Facet_Sphere_ScGeom',
 (1, 16): 'Ig2_Wall_Sphere_ScGeom',
 (7, 1): 'Ig2_Facet_Sphere_ScGeom',
 (16, 1): 'Ig2_Wall_Sphere_ScGeom'}

We can see that functors make use of symmetry (i.e. that Sphere+Wall are dispatched to the same functor as Wall+Sphere).

Finally, dispatcher can be asked to return functor suitable for given argument(s):

Yade [22]: ld=LawDispatcher([Law2_ScGeom_CpmPhys_Cpm()])

Yade [23]: ld.dispMatrix()
Out[23]: {('GenericSpheresContact', 'CpmPhys'): 'Law2_ScGeom_CpmPhys_Cpm'}

# see how the entry for ScGridCoGeom will be filled after this request
Yade [24]: ld.dispFunctor(ScGridCoGeom(),CpmPhys())
Out[24]: <Law2_ScGeom_CpmPhys_Cpm instance at 0x485f0b0>

Yade [25]: ld.dispMatrix()
{('GenericSpheresContact', 'CpmPhys'): 'Law2_ScGeom_CpmPhys_Cpm',
 ('ScGridCoGeom', 'CpmPhys'): 'Law2_ScGeom_CpmPhys_Cpm'}

OpenGL functors

OpenGL rendering is being done also by 1D functors (dispatched for the type to be rendered). Since it is sufficient to have exactly one class for each rendered type, the functors are found automatically. Their base functor types are GlShapeFunctor, GlBoundFunctor, GlIGeomFunctor and so on. These classes register the type they render using the RENDERS macro:

class Gl1_Sphere: public GlShapeFunctor {
   public :
      virtual void go(const shared_ptr<Shape>&,
         const shared_ptr<State>&,
         bool wire,
         const GLViewInfo&
      /* … */

You can list available functors of a particular type by querying child classes of the base functor:

Yade [26]: yade.system.childClasses('GlShapeFunctor')


OpenGL functors may disappear in the future, being replaced by virtual functions of each class that can be rendered.

Parallel execution

Yade was originally not designed with parallel computation in mind, but rather with maximum flexibility (for good or for bad). Parallel execution was added later; in order to not have to rewrite whole Yade from scratch, relatively non-instrusive way of parallelizing was used: OpenMP. OpenMP is standartized shared-memory parallel execution environment, where parallel sections are marked by special #pragma in the code (which means that they can compile with compiler that doesn’t support OpenMP) and a few functions to query/manipulate OpenMP runtime if necessary.

There is parallelism at 3 levels:

  • Computation, interaction (python, GUI) and rendering threads are separate. This is done via regular threads (boost::threads) and is not related to OpenMP.

  • ParallelEngine can run multiple engine groups (which are themselves run serially) in parallel; it rarely finds use in regular simulations, but it could be used for example when coupling with an independent expensive computation:

            [Engine1(),Engine2()],   # Engine1 will run before Engine2
            [Engine3()]              # Engine3() will run in parallel with the group [Engine1(),Engine2()]
                                     # arbitrary number of groups can be used

    Engine2 will be run after Engine1, but in parallel with Engine3.


    It is your reponsibility to avoid concurrent access to data when using ParallelEngine. Make sure you understand very well what the engines run in parallel do.

  • Parallelism inside Engines. Some loops over bodies or interactions are parallelized (notably InteractionLoop and NewtonIntegrator, which are treated in detail later (FIXME: link)):

    #pragma omp parallel for
    for(long id=0; id<size; id++){
       const shared_ptr<Body>& b(scene->bodies[id]);
       /* … */


    OpenMP requires loops over contiguous range of integers (OpenMP 3 also accepts containers with random-access iterators).

    If you consider running parallelized loop in your engine, always evalue its benefits. OpenMP has some overhead fo creating threads and distributing workload, which is proportionally more expensive if the loop body execution is fast. The results are highly hardware-dependent (CPU caches, RAM controller).

Maximum number of OpenMP threads is determined by the OMP_NUM_THREADS environment variable and is constant throughout the program run. Yade main program also sets this variable (before loading OpenMP libraries) if you use the -j/--threads option. It can be queried at runtime with the omp_get_max_threads function.

At places which are susceptible of being accessed concurrently from multiple threads, Yade provides some mutual exclusion mechanisms, discussed elsewhere (FIXME):


Yade provides 2 services for measuring time spent in different pars of the code. One has the granularity of engine and can be enabled at runtime. The other one is finer, but requires adjusting and recompiling the code being measured.

Per-engine timing

The coarser timing works by merely accumulating numebr of invocations and time (with the precision of the clock_gettime function) spent in each engine, which can be then post-processed by associated Python module yade.timing. There is a static bool variable controlling whether such measurements take place (disabled by default), which you can change

TimingInfo::enabled=True;            // in c++
O.timingEnabled=True                 ## in python

After running the simulation, yade.timing.stats() function will show table with the results and percentages:

Yade [27]: TriaxialTest(numberOfGrains=100).load()

Yade [28]: O.engines[0].label='firstEngine'   ## labeled engines will show by labels in the stats table

Yade [29]: import yade.timing;

Yade [30]: O.timingEnabled=True

Yade [31]: yade.timing.reset()                   ## not necessary if used for the first time

Yade [32]:; O.wait()

Yade [33]: yade.timing.stats()
Name                                                    Count                 Time            Rel. time
"firstEngine"                                        50                208us                1.88%      
InsertionSortCollider                                26               2645us               23.77%      
InteractionLoop                                      50               5219us               46.89%      
GlobalStiffnessTimeStepper                            2                 79us                0.72%      
TriaxialCompressionEngine                            50                896us                8.06%      
TriaxialStateRecorder                                 3                351us                3.16%      
NewtonIntegrator                                     50               1728us               15.53%      
TOTAL                                                                11129us              100.00%      

Exec count and time can be accessed and manipulated through Engine::timingInfo from c++ or Engine().execCount and Engine().execTime properties in Python.

In-engine and in-functor timing

Timing within engines (and functors) is based on TimingDeltas class. It is made for timing loops (functors’ loop is in their respective dispatcher) and stores cummulatively time differences between checkpoints.


Fine timing with TimingDeltas will only work if timing is enabled globally (see previous section). The code would still run, but giving zero times and exec counts.

  1. Engine::timingDeltas must point to an instance of TimingDeltas (prefferably instantiate TimingDeltas in the constructor):

    // header file
    class Law2_ScGeom_CpmPhys_Cpm: public LawFunctor {
       /* … */
          /* attrs */,
          /* constructor */
          timingDeltas=shared_ptr<TimingDeltas>(new TimingDeltas); // timingDeltas object is automatically initialized when using -DENABLE_PROFILING=1 cmake option
       // ...
  2. Inside the loop, start the timing by calling timingDeltas->start();

  3. At places of interest, call timingDeltas->checkpoint("label"). The label is used only for post-processing, data are stored based on the checkpoint position, not the label.


    Checkpoints must be always reached in the same order, otherwise the timing data will be garbage. Your code can still branch, but you have to put checkpoints to places which are in common.

    void Law2_ScGeom_CpmPhys_Cpm::go(shared_ptr<IGeom>& _geom,
                                          shared_ptr<IPhys>& _phys,
                                          Interaction* I,
                                          Scene* scene)
       timingDeltas->start();                        // the point at which the first timing starts
       // prepare some variables etc here
       // find geometrical data (deformations) here
       // compute forces here
       // apply forces, cleanup here
  4. Alternatively, you can compile Yade using -DENABLE_PROFILING=1 cmake option and use predefined macros TIMING_DELTAS_START and TIMING_DELTAS_CHECKPOINT. Without -DENABLE_PROFILING options, those macros are empty and do nothing.
    void Law2_ScGeom_CpmPhys_Cpm::go(shared_ptr<IGeom>& _geom,
                                          shared_ptr<IPhys>& _phys,
                                          Interaction* I,
                                          Scene* scene)
       // prepare some variables etc here
       // find geometrical data (deformations) here
       // compute forces here
       // apply forces, cleanup here

The output might look like this (note that functors are nested inside dispatchers and TimingDeltas inside their engine/functor):

Name                                    Count                 Time          Rel. time
ForceReseter                        400               9449μs              0.01%
BoundDispatcher                     400            1171770μs              1.15%
InsertionSortCollider               400            9433093μs              9.24%
IGeomDispatcher       400           15177607μs             14.87%
IPhysDispatcher        400            9518738μs              9.33%
LawDispatcher                       400           64810867μs             63.49%
    setup                           4926145            7649131μs             15.25%
    geom                            4926145           23216292μs             46.28%
    material                        4926145            8595686μs             17.14%
    rest                            4926145           10700007μs             21.33%
    TOTAL                                             50161117μs            100.00%
NewtonIntegrator                    400            1866816μs              1.83%
"strainer"                          400              21589μs              0.02%
"plotDataCollector"                 160              64284μs              0.06%
"damageChecker"                       9               3272μs              0.00%
TOTAL                                            102077490μs            100.00%


Do not use TimingDeltas in parallel sections, results might not be meaningful. In particular, avoid timing functors inside InteractionLoop when running with multiple OpenMP threads.

TimingDeltas data are accessible from Python as list of (label,*time*,*count*) tuples, one tuple representing each checkpoint:
deltas[0][0] # 0th checkpoint label
deltas[0][1] # 0th checkpoint time in nanoseconds
deltas[0][2] # 0th checkpoint execution count
deltas[1][0] # 1st checkpoint label
             # …

Timing overhead

The overhead of the coarser, per-engine timing, is very small. For simulations with at least several hundreds of elements, they are below the usual time variance (a few percent).

The finer TimingDeltas timing can have major performance impact and should be only used during debugging and performance-tuning phase. The parts that are file-timed will take disproportionally longer time that the rest of engine; in the output presented above, LawDispatcher takes almost ⅔ of total simulation time in average, but the number would be twice of thrice lower typically (note that each checkpoint was timed almost 5 million times in this particular case).

OpenGL Rendering

Yade provides 3d rendering based on QGLViewer. It is not meant to be full-featured rendering and post-processing, but rather a way to quickly check that scene is as intended or that simulation behaves sanely.


Although 3d rendering runs in a separate thread, it has performance impact on the computation itself, since interaction container requires mutual exclusion for interaction creation/deletion. The InteractionContainer::drawloopmutex is either held by the renderer (OpenGLRenderingEngine) or by the insertion/deletion routine.


There are 2 possible causes of crash, which are not prevented because of serious performance penalty that would result:

  1. access to BodyContainer, in particular deleting bodies from simulation; this is a rare operation, though.
  2. deleting Interaction::phys or Interaction::geom.

Renderable entities (Shape, State, Bound, IGeom, IPhys) have their associated OpenGL functors. An entity is rendered if

  1. Rendering such entities is enabled by appropriate attribute in OpenGLRenderingEngine
  2. Functor for that particular entity type is found via the dispatch mechanism.

Gl1_* functors operating on Body’s attributes (Shape, State, Bound) are called with the OpenGL context translated and rotated according to State::pos and State::ori. Interaction functors work in global coordinates.

Simulation framework

Besides the support framework mentioned in the previous section, some functionality pertaining to simulation itself is also provided.

There are special containers for storing bodies, interactions and (generalized) forces. Their internal functioning is normally opaque to the programmer, but should be understood as it can influence performance.


Scene is the object containing the whole simulation. Although multiple scenes can be present in the memory, only one of them is active. Saving and loading (serializing and deserializing) the Scene object should make the simulation run from the point where it left off.


All Engines and functors have interally a Scene* scene pointer which is updated regularly by engine/functor callers; this ensures that the current scene can be accessed from within user code.

For outside functions (such as those called from python, or static functions in Shop), you can use Omega::instance().getScene() to retrieve a shared_ptr<Scene> of the current scene.

Body container

Body container is linear storage of bodies. Each body in the simulation has its unique id, under which it must be found in the BodyContainer. Body that is not yet part of the simulation typically has id equal to invalid value Body::ID_NONE, and will have its id assigned upon insertion into the container. The requirements on BodyContainer are

  • O(1) access to elements,
  • linear-addressability (0…n indexability),
  • store shared_ptr, not objects themselves,
  • no mutual exclusion for insertion/removal (this must be assured by the called, if desired),
  • intelligent allocation of id for new bodies (tracking removed bodies),
  • easy iteration over all bodies.


Currently, there is “abstract” class BodyContainer, from which derive concrete implementations; the initial idea was the ability to select at runtime which implementation to use (to find one that performs the best for given simulation). This incurs the penalty of many virtual function calls, and will probably change in the future. All implementations of BodyContainer were removed in the meantime, except BodyVector (internally a vector<shared_ptr<Body> > plus a few methods around), which is the fastest.


Body insertion is typically used in FileGenerator‘s:

shared_ptr<Body> body(new Body);
// … (body setup)
scene->bodies->insert(body); // assigns the id

Bodies are deleted only rarely:



Since mutual exclusion is not assured, never insert/erase bodies from parallel sections, unless you explicitly assure there will be no concurrent access.


The container can be iterated over using FOREACH macro (shorthand for BOOST_FOREACH):

FOREACH(const shared_ptr<Body>& b, *scene->bodies){
   if(!b) continue;                      // skip deleted bodies
   /* do something here */

Note a few important things:

  1. Always use const shared_ptr<Body>& (const reference); that avoids incrementing and decrementing the reference count on each shared_ptr.
  2. Take care to skip NULL bodies (if(!b) continue): deleted bodies are deallocated from the container, but since body id’s must be persistent, their place is simply held by an empty shared_ptr<Body>() object, which is implicitly convertible to false.

In python, the BodyContainer wrapper also has iteration capabilities; for convenience (which is different from the c++ iterator), NULL bodies as silently skipped:

Yade [34]: O.bodies.append([Body(),Body(),Body()])
Out[34]: [0, 1, 2]

Yade [35]: O.bodies.erase(1)
Out[35]: True

Yade [36]: [ for b in O.bodies]
Out[36]: [0, 2]

In loops parallelized using OpenMP, the loop must traverse integer interval (rather than using iterators):

const long size=(long)bodies.size();       // store this value, since it doesn't change during the loop
#pragma omp parallel for
for(long _id=0; _id<size; _id++){
   const shared_ptr<Body>& b(bodies[_id]);
   if(!b) continue;
   /* … */


Interactions are stored in special container, and each interaction must be uniquely identified by pair of ids (id1,id2).

  • O(1) access to elements,
  • linear-addressability (0…n indexability),
  • store shared_ptr, not objects themselves,
  • mutual exclusion for insertion/removal,
  • easy iteration over all interactions,
  • addressing symmetry, i.e. interaction(id1,id2)≡interaction(id2,id1)


As with BodyContainer, there is “abstract” class InteractionContainer, and then its concrete implementations. Currently, only InteractionVecMap implementation is used and all the other were removed. Therefore, the abstract InteractionContainer class may disappear in the future, to avoid unnecessary virtual calls.

Further, there is a blueprint for storing interactions inside bodies, as that would give extra advantage of quickly getting all interactions of one particular body (currently, this necessitates loop over all interactions); in that case, InteractionContainer would disappear.


Creating new interactions and deleting them is delicate topic, since many eleents of simulation must be synchronized; the exact workflow is described in Handling interactions. You will almost certainly never need to insert/delete an interaction manually from the container; if you do, consider designing your code differently.

// both insertion and erase are internally protected by a mutex,
// and can be done from parallel sections safely
scene->interactions->insert(shared_ptr<Interaction>(new Interactions(id1,id2)));


As with BodyContainer, iteration over interactions should use the FOREACH macro:

FOREACH(const shared_ptr<Interaction>& i, *scene->interactions){
   if(!i->isReal()) continue;
   /* … */

Again, note the usage const reference for i. The check if(!i->isReal()) filters away interactions that exist only potentially, i.e. there is only Bound overlap of the two bodies, but not (yet) overlap of bodies themselves. The i->isReal() function is equivalent to i->geom && i->phys. Details are again explained in Handling interactions.

In some cases, such as OpenMP-loops requiring integral index (OpenMP >= 3.0 allows parallelization using random-access iterator as well), you need to iterate over interaction indices instead:

int nIntr=(int)scene->interactions->size(); // hoist container size
#pragma omp parallel for
for(int j=0; j<nIntr; j++){
   const shared_ptr<Interaction>& i=(*scene->interactions)[j];
   if(!i->isReal()) continue;
   /* … */


ForceContainer holds “generalized forces”, i.e. forces, torques, (explicit) dispalcements and rotations for each body.

During each computation step, there are typically 3 phases pertaining to forces:

  1. Resetting forces to zero (usually done by the ForceResetter engine)
  2. Incrementing forces from parallel sections (solving interactions – from LawFunctor)
  3. Reading absolute force values sequentially for each body: forces applied from different interactions are summed together to give overall force applied on that body (NewtonIntegrator, but also various other engine that read forces)

This scenario leads to special design, which allows fast parallel write access:

  • each thread has its own storage (zeroed upon request), and only writes to its own storage; this avoids concurrency issues. Each thread identifies itself by the omp_get_thread_num() function provided by the OpenMP runtime.
  • before reading absolute values, the container must be synchronized, i.e. values from all threads are summed up and stored separately. This is a relatively slow operation and we provide ForceContainer::syncCount that you might check to find cummulative number of synchronizations and compare it against number of steps. Ideally, ForceContainer is only synchronized once at each step.
  • the container is resized whenever an element outside the current range is read/written to (the read returns zero in that case); this avoids the necessity of tracking number of bodies, but also is potential danger (such as scene->forces.getForce(1000000000), which will probably exhaust your RAM). Unlike c++, Python does check given id against number of bodies.
// resetting forces (inside ForceResetter)

// in a parallel section
scene->forces.addForce(id,force); // add force

// container is not synced after we wrote to it, sync before reading
const Vector3r& f=scene->forces.getForce(id);

Synchronization is handled automatically if values are read from python:

Yade [37]: O.bodies.append(Body())
Out[37]: 3

Yade [38]: O.forces.addF(0,Vector3(1,2,3))
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 O.forces.addF(0,Vector3(1,2,3))

NameError: name 'Vector3' is not defined

Yade [39]: O.forces.f(0)
Out[39]: Vector3(0,0,0)

Yade [40]: O.forces.f(100)
IndexError                                Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 O.forces.f(100)

IndexError: Body id out of range.

Handling interactions

Creating and removing interactions is a rather delicate topic and number of components must cooperate so that the whole behaves as expected.

Terminologically, we distinguish

potential interactions,
having neither geometry nor physics. Interaction.isReal can be used to query the status (Interaction::isReal() in c++).
real interactions,
having both geometry and physics. Below, we shall discuss the possibility of interactions that only have geometry but no physics.

During each step in the simulation, the following operations are performed on interactions in a typical simulation:

  1. Collider creates potential interactions based on spatial proximity. Not all pairs of bodies are susceptible of entering interaction; the decision is done in Collider::mayCollide:

    • clumps may not enter interactions (only their members can)
    • clump members may not interact if they belong to the same clump
    • bitwise AND on both bodies’ masks must be non-zero (i.e. there must be at least one bit set in common)
  2. Collider erases interactions that were requested for being erased (see below).

  3. InteractionLoop (via IGeomDispatcher) calls appropriate IGeomFunctor based on Shape combination of both bodies, if such functor exists. For real interactions, the functor updates associated IGeom. For potential interactions, the functor returns


    if there is no geometrical overlap, and the interaction will stillremain potential-only


    if there is geometrical overlap; the functor will have created an IGeom in such case.


    For real interactions, the functor must return true, even if there is no more spatial overlap between bodies. If you wish to delete an interaction without geometrical overlap, you have to do this in the LawFunctor.

    This behavior is deliberate, since different laws have different requirements, though ideally using relatively small number of generally useful geometry functors.


    If there is no functor suitable to handle given combination of shapes, the interaction will be left in potential state, without raising any error.

  4. For real interactions (already existing or just created in last step), InteractionLoop (via IPhysDispatcher) calls appropriate IPhysFunctor based on Material combination of both bodies. The functor must update (or create, if it doesn’t exist yet) associated IPhys instance. It is an error if no suitable functor is found, and an exception will be thrown.

  5. For real interactions, InteractionLoop (via LawDispatcher) calls appropriate LawFunctor based on combination of IGeom and IPhys of the interaction. Again, it is an error if no functor capable of handling it is found.

  6. LawDispatcher takes care of erasing those interactions that are no longer active (such as if bodies get too far apart for non-cohesive laws; or in case of complete damage for damage models). This is triggered by the LawFunctor returning false. For this reason it is of upmost importance for the LawFunctor to return consistently.

Such interaction will not be deleted immediately, but will be reset to potential state. At the next execution of the collider InteractionContainer::conditionalyEraseNonReal will be called, which will completely erase interactions only if the bounding boxes ceased to overlap; the rest will be kept in potential state.

Creating interactions explicitly

Interactions may still be created explicitly with utils.createInteraction, without any spatial requirements. This function searches current engines for dispatchers and uses them. IGeomFunctor is called with the force parameter, obliging it to return true even if there is no spatial overlap.

Associating Material and State types

Some models keep extra state information in the Body.state object, therefore requiring strict association of a Material with a certain State (for instance, CpmMat is associated to CpmState and this combination is supposed by engines such as CpmStateUpdater).

If a Material has such a requirement, it must override 2 virtual methods:

  1. Material.newAssocState, which returns a new State object of the corresponding type. The default implementation returns State itself.
  2. Material.stateTypeOk, which checks whether a given State object is of the corresponding type (this check is run at the beginning of the simulation for all particles).

In c++, the code looks like this (for CpmMat):

class CpmMat: public FrictMat {
      virtual shared_ptr<State> newAssocState() const { return shared_ptr<State>(new CpmState); }
      virtual bool stateTypeOk(State* s) const { return (bool)dynamic_cast<CpmState*>(s); }
   /* ... */

This allows one to construct Body objects from functions such as utils.sphere only by knowing the requires Material type, enforcing the expectation of the model implementor.

Runtime structure

Startup sequence

Yade’s main program is python script in core/main/; the build system replaces a few ${variables} in that file before copying it to its install location. It does the following:

  1. Process command-line options, set environment variables based on those options.
  2. Import main yade module (import yade), residing in py/ This module locates plugins (recursive search for files lib*.so in the lib installation directory). yade.boot module is used to setup temporary directory, … and, most importantly, loads plugins.
  3. Manage further actions, such as running scripts given at command line, opening qt.Controller (if desired), launching the ipython prompt.


There are several “global variables” that are always accessible from c++ code; properly speaking, they are Singletons, classes of which exactly one instance always exists. The interest is to have some general functionality acessible from anywhere in the code, without the necessity of passing pointers to such objects everywhere. The instance is created at startup and can be always retrieved (as non-const reference) using the instance() static method (e.g. Omega::instance().getScene()).

There are 3 singletons:

Handles serialization/deserialization; it is not used anywhere except for the serialization code proper.
Registers classes from plugins and able to factor instance of a class given its name as string (the class must derive from Factorable). Not exposed to python.
Access to simulation(s); deserves separate section due to its importance.


The Omega class handles all simulation-related functionality: loading/saving, running, pausing.

In python, the wrapper class to the singleton is instantiated [5] as global variable O. For convenience, Omega is used as proxy for scene’s attribute: although multiple Scene objects may be instantiated in c++, it is always the current scene that Omega represents.

The correspondence of data is literal: Omega.materials corresponds to Scene::materials of the current scene; likewise for materials, bodies, interactions, tags, cell, engines, initializers, miscParams.

To give an overview of (some) variables:

Python c++
Omega.iter Scene::iter
Omega.dt Scene::dt
Omega.time Scene::time
Omega.realtime Omega::getRealTime()
Omega.stopAtIter Scene::stopAtIter

Omega in c++ contains pointer to the current scene (Omega::scene, retrieved by Omega::instance().getScene()). Using Omega.switchScene, it is possible to swap this pointer with Omega::sceneAnother, a completely independent simulation. This can be useful for example (and this motivated this functionality) if while constructing simulation, another simulation has to be run to dynamically generate (i.e. by running simulation) packing of spheres.


[5]It is understood that instantiating Omega() in python only instantiates the wrapper class, not the singleton itself.

Engine loop

Running simulation consists in looping over Engines and calling them in sequence. This loop is defined in Scene::moveToNextTimeStep function in core/Scene.cpp. Before the loop starts, O.initializers are called; they are only run once. The engine loop does the following in each iteration over O.engines:

  1. set Engine::scene pointer to point to the current Scene.
  2. Call Engine::isActivated(); if it returns false, the engine is skipped.
  3. Call Engine::action()
  4. If O.timingEnabled, increment Engine::execTime by the difference from the last time reading (either after the previous engine was run, or immediately before the loop started, if this engine comes first). Increment Engine::execCount by 1.

After engines are processed, virtual time is incremented by timestep and iteration number is incremented by 1.

Background execution

The engine loop is (normally) executed in background thread (handled by SimulationFlow class), leaving foreground thread free to manage user interaction or running python script. The background thread is managed by and O.pause() commands. Foreground thread can be blocked until the loop finishes using O.wait().

Single iteration can be run without spawning additional thread using O.step().

Python framework

Wrapping c++ classes

Each class deriving from Serializable is automatically exposed to python, with access to its (registered) attributes. This is achieved via YADE_CLASS_BASE_DOC_* macro family. All classes registered in class factory are default-constructed in Omega::buildDynlibDatabase. Then, each serializable class calls Serializable::pyRegisterClass virtual method, which injects the class wrapper into (initially empty) yade.wrapper module. pyRegisterClass is defined by YADE_CLASS_BASE_DOC and knows about class, base class, docstring, attributes, which subsequently all appear in boost::python class definition.

Wrapped classes define special constructor taking keyword arguments corresponding to class attributes; therefore, it is the same to write:

Yade [41]: f1=ForceEngine()

Yade [42]: f1.ids=[0,4,5]

Yade [43]: f1.force=Vector3(0,-1,-2)
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 f1.force=Vector3(0,-1,-2)

NameError: name 'Vector3' is not defined


Yade [44]: f2=ForceEngine(ids=[0,4,5],force=Vector3(0,-1,-2))
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 f2=ForceEngine(ids=[0,4,5],force=Vector3(0,-1,-2))

NameError: name 'Vector3' is not defined

Yade [45]: print f1.dict()
{'ompThreads': -1, 'force': Vector3(0,0,0), 'ids': [0, 4, 5], 'dead': False, 'label': ''}

Yade [46]: print f2.dict()
NameError                                 Traceback (most recent call last)
/home/buildbot/yade/yade-full/install/lib/x86_64-linux-gnu/yade-buildbot/py/yade/__init__.pyc in <module>()
----> 1 print f2.dict()

NameError: name 'f2' is not defined

Wrapped classes also inherit from Serializable several special virtual methods: dict() returning all registered class attributes as dictionary (shown above), clone() returning copy of instance (by copying attribute values), updateAttrs() and updateExistingAttrs() assigning attributes from given dictionary (the former thrown for unknown attribute, the latter doesn’t).

Read-only property name wraps c++ method getClassName() returning class name as string. (Since c++ class and the wrapper class always have the same name, getting python type using __class__ and its property __name__ will give the same value).

Yade [47]: s=Sphere()

Yade [48]: s.__class__.__name__
Out[48]: 'Sphere'

Subclassing c++ types in python

In some (rare) cases, it can be useful to derive new class from wrapped c++ type in pure python. This is done in the yade.pack module module: Predicate is c++ base class; from this class, several c++ classes are derived (such as inGtsSurface), but also python classes (such as the trivial inSpace predicate). inSpace derives from python class Predicate; it is, however, not direct wrapper of the c++ Predicate class, since virtual methods would not work.

boost::python provides special boost::python::wrapper template for such cases, where each overridable virtual method has to be declared explicitly, requesting python override of that method, if present. See Overridable virtual functions for more details.

Reference counting

Python internally uses reference counting on all its objects, which is not visible to casual user. It has to be handled explicitly if using pure Python/C API with Py_INCREF and similar functions.

boost::python used in Yade fortunately handles reference counting internally. Additionally, it automatically integrates reference counting for shared_ptr and python objects, if class A is wrapped as boost::python::class_<A,shared_ptr<A>>. Since all Yade classes wrapped using YADE_CLASS_BASE_DOC_* macro family are wrapped in this way, returning shared_ptr<…> objects from is the preferred way of passing objects from c++ to python.

Returning shared_ptr is much more efficient, since only one pointer is returned and reference count internally incremented. Modifying the object from python will modify the (same) object in c++ and vice versa. It also makes sure that the c++ object will not be deleted as long as it is used somewhere in python, preventing (important) source of crashes.

Custom converters

When an object is passed from c++ to python or vice versa, then either

  1. the type is basic type which is transparently passed between c++ and python (int, bool, std::string etc)
  2. the type is wrapped by boost::python (such as Yade classes, Vector3 and so on), in which case wrapped object is returned; [6]

Other classes, including template containers such as std::vector must have their custom converters written separately. Some of them are provided in py/wrapper/customConverters.cpp, notably converters between python (homogeneous, i.e. with all elements of the same type) sequences and c++ std::vector of corresponding type; look in that source file to add your own converter or for inspiration.

When an object is crossing c++/python boundary, boost::python’s global “converters registry” is searched for class that can perform conversion between corresponding c++ and python types. The “converters registry” is common for the whole program instance: there is no need to register converters in each script (by importing _customConverters, for instance), as that is done by yade at startup already.


Custom converters only work for value that are passed by value to python (not “by reference”): some attributes defined using YADE_CLASS_BASE_DOC_* macro family are passed by value, but if you define your own, make sure that you read and understand Why is my automatic to-python conversion not being found?.

In short, the default for def_readwrite and def_readonly is to return references to underlying c++ objects, which avoids performing conversion on them. For that reason, return value policy must be set to return_by_value explicitly, using slighly more complicated add_property syntax, as explained at the page referenced.

[6]Wrapped classes are automatically registered when the class wrapper is created. If wrapped class derives from another wrapped class (and if this dependency is declared with the boost::python::bases template, which Yade’s classes do automatically), parent class must be registered before derived class, however. (This is handled via loop in Omega::buildDynlibDatabase, which reiterates over classes, skipping failures, until they all successfully register) Math classes (Vector3, Matrix3, Quaternion) are wrapped in minieigen, which is available as a separate package. Use your package manager to install it.

Maintaining compatibility

In Yade development, we identified compatibility to be very strong desire of users. Compatibility concerns python scripts, not simulations saved in XML or old c++ code.

Renaming class

Script scripts/ should be used to rename class in c++ code. It takes 2 parameters (old name and new name) and must be run from top-level source directory:

$ scripts/ OldClassName NewClassName
Replaced 4 occurences, moved 0 files and 0 directories
Update python scripts (if wanted) by running: perl -pi -e 's/\bOldClassName\b/NewClassName/g' `ls **/*.py |grep -v py/`

This has the following effects:

  1. If file or directory has basename OldClassName (plus extension), it will be renamed using bzr.

  2. All occurences of whole word OldClassName will be replaced by NewClassName in c++ sources.

  3. An extry is added to py/, which contains map of deprecated class names. At yade startup, proxy class with OldClassName will be created, which issues a DeprecationWarning when being instantiated, informing you of the new name you should use; it creates an instance of NewClassName, hence not disruting your script’s functioning:

    Yade [3]: SimpleViscoelasticMat()
    /usr/local/lib/yade-trunk/py/yade/ DeprecationWarning: Class `SimpleViscoelasticMat' was renamed to (or replaced by) `ViscElMat', update your code! (you can run 'yade --update' to do that automatically)
    ->  [3]: <ViscElMat instance at 0x2d06770>

As you have just been informed, you can run yade --update to all old names with their new names in scripts you provide:

$ yade-trunk --update some/where/

This gives you enough freedom to make your class name descriptive and intuitive.

Renaming class attribute

Renaming class attribute is handled from c++ code. You have the choice of merely warning at accessing old attribute (giving the new name), or of throwing exception in addition, both with provided explanation. See deprec parameter to YADE_CLASS_BASE_DOC_* macro family for details.

Debian packaging instructions

In order to make parallel installation of several Yade version possible, we adopted similar strategy as e.g. gcc packagers in Debian did:

  1. Real Yade packages are named yade-0.30 (for stable versions) or yade-bzr2341 (for snapshots).
  2. They provide yade or yade-snapshot virtual packages respectively.
  3. Each source package creates several installable packages (using bzr2341 as example version):
    1. yade-bzr2341 with the optimized binaries; the executable binary is yade-bzr2341 (yade-bzr2341-multi, …)
    2. yade-bzr2341-dbg with debug binaries (debugging symbols, non-optimized, and with crash handlers); the executable binary is yade-bzr2341-dbg
    3. yade-bzr2341-doc with sample scripts and some documentation (see bug #398176 however)
    4. (future?) yade-bzr2341-reference with reference documentation (see bug #401004)
  4. Using Debian alternatives, the highest installed package provides additionally commands without the version specification like yade, yade-multi, … as aliases to that version’s binaries. (yade-dbg, … for the debuggin packages). The exact rule is:
    1. Stable releases have always higher priority than snapshots
    2. Higher versions/revisions have higher pripority than lower versions/revisions.

Prepare source package

Debian packaging files are located in debian/ directory. They contain build recipe debian/rules, dependecy and package declarations debian/control and maintainer scripts. Some of those files are only provided as templates, where some variables (such as version number) are replaced by special script.

The script scripts/debian-prep processes templates in debian/ and creates files which can be used by debian packaging system. Before running this script:

  1. If you are releasing stable version, make sure there is file named RELEASE containing single line with version number (such as 0.30). This will make scripts/debian-prep create release packages. In absence of this file, snapshots packaging will be created instead. Release or revision number (as detected by running bzr revno in the source tree) is stored in VERSION file, where it is picked up during package build and embedded in the binary.
  2. Find out for which debian/ubuntu series your package will be built. This is the name that will appear on the top of (newly created) debian/changelog file. This name will be usually unstable, testing or stable for debian and karmic, lucid etc for ubuntu. WHen package is uploaded to Launchpad’s build service, the package will be built for this specified release.

Then run the script from the top-level directory, giving series name as its first (only) argument:

$ scripts/debian-prep lucid

After this, signed debian source package can be created:

$ debuild -S -sa -k62A21250 -I -Iattic

(-k gives GPG key identifier, -I skips .bzr and similar directories, -Iattic will skip the useless attic directory).

Create binary package

Local in-tree build
Once files in debian/ are prepared, packages can be build by issuing:: $ fakeroot debian/rules binary
Clean system build
Using pbuilder system, package can be built in a chroot containing clean debian/ubuntu system, as if freshly installed. Package dependencies are automatically installed and package build attempted. This is a good way of testing packaging before having the package built remotely at Launchpad. Details are provided at wiki page.
Launchpad build service

Launchpad provides service to compile package for different ubuntu releases (series), for all supported architectures, and host archive of those packages for download via APT. Having appropriate permissions at Launchpad (verified GPG key), source package can be uploaded to yade’s archive by:

$ dput ppa:yade-users/ppa ../yade-bzr2341_1_source.changes

After several hours (depending on load of Launchpad build farm), new binary packages will be published at

This process is well documented at

Table Of Contents

Previous topic

User’s manual

Next topic


This Page