Developer Guide

This page presents instructions on how to develop code using IMP. Developers who wish to contribute code back to IMP or distribute their code should also read the Contributing code to IMP page.

1. Getting around

The input files in the IMP directory are structured as follows:

tools contains various command line utilities for use by developers. They are documented below.
doc contains inputs for general IMP overview documentation (such as this page), as well as configuration scripts for doxygen.
applications contains various applications implementing using a variety of IMP modules.
kernel and the subdirectories of module/ each defines a module and have the same structure. The directory for module name has the following structure
- include contains the C++ header files
- src contains the C++ source files
- bin contains C++ source files each of which is built into an executable
- pyext contains files defining the python interface to the module as well as python source files (in pyext/src)
- test contains test files. When scons test or scons name-test is run each file in this directory named test_ is executed (after being built if it is a .cpp file)
- doc contains the overview documentation for the file (in the SConscript file) as well as any other documentation that is provided via .dox files
- examples contains examples, in python as well as any data needed for examples
- data contains any data files needed by the module

When IMP is built, the build directory is created and filled with the results of the build. build contains a number of subdirectories. They are

include which includes all the headers. The headers for the kernel are placed in include/IMP and those for module name are placed in include/IMP/name
lib where the C++ and python libraries are placed. Module name is built into a C++ library lib/libimp_name.so (or .dylib on a mac) and a python library with python files located in lib/IMP/name and the binary part in lib/_IMP_name.so.
doc where the html documentation is placed in doc/html and the examples in doc/examples with a subdirectory for each module
data where each module gets a subdirectory for its data.

Unfortunately, various intermediate files from the build are scattered throughout the module and kernel hierarchies. This messiness is part of the reason we strongly recommend doing an out of source build.

When IMP is installed, the structure from the build directory is moved over more or less intact except that the C++ and python libraries are put in the (different) appropriate locations.

Writing new functions and classes

The easiest way to start writing new functions and classes is to create a new module using the make-module script. This creates a new module in the modules directory, complete with example code.

We highly recommend using a revision control system such as Subversion or GIT to keep track of changes to your module.

If, instead, you choose to add code to an existing module you need to consult with the person who people who control that module. Their names can be found on the module main page.

When designing the interface for your new code, you should

search IMP for similar functionality and, if there is any, adapt the existing interface for your purposes. For example, the existing IMP::atom::read_pdb() and IMP::atom::write_pdb() functions provide templates that should be used for the design of any functions that create particles from a file or write particles to a file. Since IMP::atom::BondDecorator, IMP::algebra::Segment3D and IMP::display::Geometry all use methods like IMP::algebra::Segment3D::get_point() to access the endpoints of a segment, any new object which defines similar point-based geometry should do likewise.

think about how other people are likely to use the code. For example, not all molecular hierarchies have atoms as their leaves, so make sure your code searches for arbitrary IMP::core::XYZDecorator particles rather than atoms if you only care about the geometry.

look for easy ways of splitting the functionality into pieces. It generally makes sense, for example, to split selection of the particles from the action taken on them, either by accepting a IMP::ParticleRefiner, or a IMP::SingletonContainer or just an arbitrary IMP::Particles object.

You may want to read the design example for some suggestions on how to go about implementing your functionality in IMP.

Debugging and testing your code

Ensuring that your code is correct can be very difficult, so IMP provides a number of tools to help you out.

The first set are assert-style macros:

IMP_USAGE_CHECK() which should be used to check that arguments to functions and methods satisfy the preconditions.

IMP_INTERNAL_CHECK() which should be used to verify internal state and return values to make sure they satisfy pre and post-conditions.

See Error reporting/checking page for more details. As a general guideline, any improper usage to produce at least a warning all return values should be checked by such code.

The second is logging macros such as:

IMP_LOG() which allows controlled display of messages about what the code is doing. See logging for more information.

Finally, each module has a set of unit tests. These are python scripts which test the behavior of a particular piece of the modules API. The scripts are located in the modules/modulename/test directory. These tests should try, as much as possible to provide independent verification of the correctness of the C++ code. The command

 scons test

 scons modulename-test

run all modules unit tests or only those for the module modulename, respectively. Any file in that directory or a subdirectory whose name matches test_*.py is considered a test. The python files are scanned for classes which inherit from IMP.test.TestCase. For each such class found, any method whose name starts with test_ is run.

Some tests will require input files or temporary files. Input files should be placed in a directory called input in the test directory. The test script should then call

 self.get_input_file_name(file_name)

to get the true path to the file. Likewise, appropriate names for temporary files should be found by calling

 self.get_tmp_file_name(file_name)

. Temporary files will be located in build/tmp. The test should remove temporary files after using them.

Coding conventions

Make sure you read the API conventions page first.

To ensure code consistency and readability, certain conventions must be adhered to when writing code for IMP. Some of these conventions are automatically checked for by source control before allowing a new commit, and can also be checked yourself in new code by running

 scons standards

Indentation

All C++ headers and code should be indented in 'Linux' style, with 2-space indents. Do not use tabs. This is roughly the output of Artistic Style run like

 astyle --convert-tabs --style=linux --indent=spaces=2 --unpad=paren --pad=oper

. Split lines if necessary to ensure that no line is longer than 80 characters.

Rationale: Different users have different-sized windows or terminals, and different tab settings, but everybody can read 80 column output without tabs.

All Python code should conform to the Python style guide. In essence this translates to 4-space indents, no tabs, and similar class, method and variable naming to the C++ code. You can ensure that your Python code is correctly indented by using the

 tools/reindent.py

script, available as part of the IMP distribution.

Names

See the names section of the IMP conventions page. In addition, developers should be aware that

all preprocessor symbols (things created by #define) must begin with IMP and no IMP code should depend on preprocessor symbols which do not start with IMP.
names of files that implement a single class should be named for that class; for example the SpecialVector class could be implemented in SpecialVector.h and SpecialVector.cpp
files that provide free functions or macros should be given names separated_by_underscores, for example container_macros.h
Functions which take a parameter which has units should have the unit as part of the function name, for example IMP::atom::SimulationParameters::set_maximum_time_step_in_femtoseconds(). Remember the Mars orbiter. The exception to this is distance and force numbers which should always be in angstroms and kcal/mol angstrom respectively unless otherwise stated.

Rationale: This makes it easier to tell between class names and function names where this is ambiguous (particularly an issue with the Python interface). The Python guys also mandate CamelCase for their class names, so this avoids any need to rename classes between C++ and Python to ensure clean Python code. Good naming is especially important with preprocessor symbols since these have file scope and so can change the meaning of other people's code.

Passing and storing data

When a class or function takes a set of particles which are expected to be those of a particular type of decorator, it should take a list of decorators instead. eg IMP::core::transform() takes a IMP::core::XYZ. This makes it clearer what attributes the particle is required to have as well as allows functions to be overloaded (so there can be an IMP::core::transform() which takes IMP::core::RigidBody particles instead).

IMP::Restraint and IMP::ScoreState classes should generally use a IMP::SingletonContainer (or other type of Container) to store the set of IMP::Particle objects that they act on.

Store collections of IMP::Object-derived or IMP::Decorator-derived objects of type Name using a Names. Declare functions that accept them to take a NamesTemp (Names is a NamesTemp). Names are reference counted (see IMP::RefCounted for details), NamesTemp are not.

Display

All classes must have a show method which takes an optional std::ostream and prints information about the object (see IMP::Object::show() for an example). The helper macros, such as IMP_RESTRAINT() define such a method. In addition they must have operator<< defined. This can be easily done using the IMP_OUTPUT_OPERATOR() macro once the show method is defined. Note that operator<< writes human readable information. Add a write method if you want to provide output that can be read back in.

Errors

Classes and methods should use IMP exceptions to report errors. See IMP::Exception for a list of existing exceptions. See assert for a list of functions to aid in error reporting and detection.

Namespaces

Use the provided IMPMODULE_BEGIN_NAMESPACE, IMPMODULE_END_NAMESPACE, IMPMODULE_BEGIN_INTERNAL_NAMESPACE and IMPMODULE_END_INTERNAL_NAMESPACE macros to put declarations in a namespace appropriate for module MODULE.

Each module has an internal namespace, module_name::internal and an internal include directory modulename/internal. Any function which is

not intended to be part of the API,
not documented,
liable to change without notice,
or not tested

should be declared in an internal header and placed in the internal namespace.

The functionality in such internal headers is

not exported to python
and not part of of documented API

As a result, such functions do not need to obey all the coding conventions (but we recommend that they do).

Documenting your code

IMP is documented using Doxygen. See documenting source code with doxygen to get started. We use //! and /** ... * / blocks for documentation.

Python code should provide Python doc strings.

All headers not in internal directories are parsed through Doxygen. Any function that you do not want documented (for example, because it is not well tested), hide by surrounding with

#ifndef IMP_DOXYGEN
void messy_poorly_thought_out_function();
#endif

We provide a number of extra Doxygen commands to aid in producing nice IMP documentation. The commands are used by writing \commandname{args} or \commandname if there are no arguments.

When you want to specify some command-line command do
```
    \command{the command text}
```
which produces
```
 the command text 
```

To produce a link to a page on the Sali lab web site do
```
    \salilab{imp, the IMP project}
```
which produces the IMP project

To produce a link to the outside world do
```
    \external{boost.org, Boost}
```
produces Boost

When writing the name IMP do
```
    \imp
```
so that no link is produced (IMP as opposed to IMP).

Sections of documentation that are only for people developing IMP code should be marked with
```
  \advanceddoc You can tweak this class in various ways in order to optimize its
  performance. 
```
Similarly advanced methods should be marked with
```
  \advancedmethod
```
To produce
Note:
Familiarity with these methods is not required to use imp..
General warning messages can be produced using
```
  \warning Be afraid, be very afraid.
```
which produces
Warning: Be afraid, be very afraid.
To mark that some part of the API has not yet been well planned at may change using \unstable{Classname}. The documentation will include a disclaimer and the class or function will be added to a list of unstable classes. It is better to simply hide such things from Doxygen.

To mark a method as not having been well tested yet, use \untested{Classname}.

To mark a method as not having been implemented, use \untested{Classname}.

To note that a class supports comparisons (eg <, >, ==, != etc) use \comparable and then hide the comparison functions from Doxygen (there are a lot of them and they aren't very interesting).

Useful Scripts

IMP provides a variety of scripts to aid the lives of developers.

Generate SConscripts

The SConscripts in a number of the modules list all of the header and cpp files which are part of the module (those of other modules automatically build this list at compile time). These lists can be generated using the make-sconscripts script. To run it to rebuild the SConscripts for the module modulename do

 ./tools/make-sconscripts modulename

Making a module

Creating such a module is the easiest way to get started developing code for IMP. First, choose a name for the module. The name should only contain letters, numbers and underscores as it needs to be a valid file name as well as an identifier in Python and C++.

To create the module do

 ./tools/make-module my_module

Then, if you run scons with localmodules=True, your new module will be built. The new module includes a number of examples and comments to help you add code to the module.

You can use your new module in a variety of ways:

add C++ code to your module by putting .h files in modules/my_module/include and .cpp files in modules/my_module/src. In order to use use your new functions and classes in python, you must add a line include "IMP/my_module/myheader.h" near the end of the file modules/my_module/pyext/my_module.i.
write C++ programs using IMP by creating .cpp files in modules/my_module/bin. Each .cpp file placed there is built into a separate executable.
add python code to your library by putting a .py file in modules/my_module/pyext/my_module/
add python code to your library by by adding pythoncode blocks to modules/my_module/pyext/my_module.i.
add test code to your library by putting .py files in modules/my_module/test or a subdirectory.

If you feel your module is of interest to other IMP users and developers, see the contributing code to IMP section.

If you document your code, running

 scons doc

will build documentation of all of the modules including yours. To access the documentation, open doc/html/index.html.

Contributing code back to the repository

In order to be shared with others as part of the IMP distribution, code needs to be of higher quality and more thoroughly vetted than typical research code. As a result, it may make sense to keep the code as part of a private module until you better understand what capabilities can be cleanly offered to others.

The first set of questions to answer are

What exactly is the functionality I would like to contribute? Is it a single function, a single Restraint, a set of related classes and functions?

Is there similar functionality already in IMP? If so, it might make more sense to modify the existing code in cooperation with its author. At the very least, the new code needs to respect the conventions established by the prior code in order to maintain consistency.

Where should the new functionality go? It can either be added to an existing module or as part of a new module. If adding to an existing module, you must communicate with the authors of that module to get permission and coordinate changes.

Should the functionality be written in C++ or Python? In general, we suggest C++ if you are comfortable programming in that language as that makes the functionality available to more people. See Python/C++ differences for more considerations.

You are encouraged to post to the imp-dev to find help answering these questions as it can be hard to grasp all the various pieces of functionality already in the repository.

All code contributed to IMP

must follow the IMP coding conventions
should follow general good C++ programming practices
must have unit tests
must pass all unit tests
must have documentation
must build on all supported compilers (roughly, recent versions of gcc and Visual C++) without warnings
should have examples
must not have warnings when Doxygen is run (scons doc)

The next suggestions provide more details about the above choices and how to implement them.

Submitting to a module

Small pieces of functionality or extensions to existing functionality probably should be submitted to an existing module. Please contact the authors of the appropriate module and discuss the submission and how the code will be maintained.

A list of all current modules in the IMP repository can be found in the modules list or from the modules tab at the top of this page.

As always, if in doubt, post to imp-dev.

Patches to modules for which you have write access can be submitted directly by doing:

 svn commit -m "message describing the patch" files or directories to submit

Submitting to a module

If you have a large group of related functionality to submit, it may make sense to create a new module in svn. Please post to imp-dev to discuss your plans.

Once you have submitted code

Once you have submitted code, you should monitor the Nightly build status to make sure that your code builds on all platforms and passes the unit tests. Please fix all build problems as fast as possible.

The following sorts of changes must be announced on the imp-dev mailing list before being made

changes to existing kernel or core APIs
significant additions to kernel or core

We recommend that changes be posted to the list a day or two before they are made so as to give everyone adequate time to comment.

In addition to monitoring the imp-dev list, developers who have a module or are committing patches to svn may want to subscribe to the imp-commits email list which receives notices of all changes made to the IMP SVN repository.

Good programming practices

The contents of this page are aimed at C++ programmers, but most apply also to python.

General resources

Two excellent sources for general C++ coding guidelines are

C++ Coding Standards by Sutter and Alexandrescu

Effective C++ by Meyers

IMP endeavors to follow all the of the guidelines published in those books. The Sali lab owns copies of both of these books that you are free to borrow.

IMP gotchas

Below are a suggestions prompted by bugs found in code submitted to IMP.

Never use 'using namespace' outside of a function; instead explicitly provide the namespace. (This avoids namespace pollution, and removes any ambiguity.)

Never use the preprocessor to define constants. Use const variables instead. Preprocessor symbols don't have scope or type and so can have unexpected effects.

Pass other objects by value or by const & (if the object is large) and store copies of them.

Never expose member variables in an object which has methods. All such member variables should be private.

Don't derive a class from another class simply to reuse some code that the base class provides - only do so if your derived class could make sense when cast to the base class. As above, reuse existing code by pulling it into a function.

Clearly mark any file that is created by a script so that other people know to edit the original file.

Always return a const value or const ref if you are not providing write access. Returning a const copy means the compiler will report an error if the caller tries to modify the return value without creating a copy of it.

Include files from the local module first, then files from the other IMP modules and kernel and finally outside includes. This makes any dependencies in your code obvious, and by including standard headers after IMP headers, any missing includes in the headers themselves show up early (rather than being masked by other headers you include).

    #include <IMP/mymodule/mymodule_exports.h>
    #include <IMP/mymodlule/MyRestraint.h>
    #include <IMP/Restraint.h>
    #include <vector>

Use double variables for all computational intermediates.

Avoid using nested classes in the API as SWIG can't wrap them properly. If you must use use nested classes, you will have to do more work to provide a Python interface to your code.

Delay initialization of keys until they are actually needed (since all initialized keys take up memory within each particle, more or less). The best way to do this is to have them be static variables in a static function:
```
      FloatKey get_my_float_key() {
               static FloatKey k("hello");
               return k;
      }
```

One is the almost always the right number:
- Information should be stored in exactly one place. Duplicated information easily gets out of sync.
- A given piece of code should only appear once. Do not copy, paste and modify to create new functionality. Instead, figure out a way to reuse the existing code by pulling it into an internal function and adding extra parameters. If you don't, when you find bugs, you won't remember to fix them in all the copies of the code.
- There should be exactly one way to represent any particular state. If there is more than one way, anyone who writes library code which uses that type of state has to handle all ways. For example, there is only one scheme for representing proteins, namely the IMP::atom::MolecularHierarchyDecorator.
- Each class/method should do exactly one thing. The presence of arguments which dramatically change the behavior of the class/method is a sign that it should be split. Splitting it can make the code simpler, expose the common code for others to use and make it harder to make mistakes by getting the mode flag wrong.
- Methods should take at most one argument of each type (and ideally only one argument). If there are several arguments of the same types (eg two different double parameters) it is easy for a user to mix up the order of arguments and the compiler will not complain. int and double count as equivalent types for this rule since the compiler will transparent convert an int into a double.