ParaMonte Fortran development guidelines | ParaMonte: Parallel Monte Carlo and Machine Learning Library

Edit this page

Thank you for your interest in contributing to ParaMonte! Your help is very much appreciated. Below are some tips and guidelines to get started with contributing to the ParaMonte Fortran routines.

Initial steps

First, read the general development guidelines.
Take a look at the issues page. Make sure you find an open issue about the Fortran routines and that you do not duplicate someone else’s work.
If your contribution does not exist as an issue, post a new issue discussing the changes you’re proposing to implement, whether bug fix(es) or enhancement/feature request(s), or give the rest of developers a heads up that you are going to start work on an open issue.
Fork the ParaMonte git project to your private account.

Kernel development conventions

Pay careful attention to the following conventions used in the development of the Fortran routines.

Preprocessor directives

The Fortran routines of ParaMonte heavily rely on the compiler preprocessing directives to implement multiple completely different parallelism paradigms (serial, MPI, Coarray) that are specifically tailored for the needs of multiple completely different programming languages (C, C++, Fortran, Julia, MATLAB, Python, R, …) for different builds (debug, testing, release), for different operating systems (Windows, Linux, macOS, WSL), and many more.

Despite significant efforts to minimize the use of preprocessor directive in the Fortran routines, their frequent usage is unavoidable as they greatly minimize the cost of development. Simultaneously, however, the preprocessor directive can increase confusion and and lead to implicit bugs that are hard to detect.

Warning: Any code section that requires some specific operating system, platform, compiler, programming language, compiler settings, or parallelism paradigms must be fenced with the appropriate preprocessor macros such that it is executed only in the appropriate circumstances. Failure to do so, will lead to a non-portable codebase. In general, there must exist an #else preprocessor section for any fenced code section that appropriately handles other possible scenarios, if the other scenarios are not possible, it is followed by an #error "error message" preprocessor directive.

Important: The preprocessor macros must be always defined with uppercase characters. This is to easily distinguish macros from other variables in the regular sections of the code.

Important: Define macros as propositions that evaluate to either TRUE or FALSE. For example, WINDOWS_ENABLED or MPI_ENABLED.

Tip: When developing or debugging a particular feature that is bound to a specific set of preprocessor macros, you can request the compiler to generate intermediate preprocessed files that are free from preprocessor directives and let you focus on the problem better. In doing so, keep in mind that every change you make to the intermediate codes has be carefully implemented in the original set of codes.

Platform preprocessor directives

If there is any section of the code that must be exclusively executed on a particular platform or Operating System (OS), it must be properly fenced by the appropriate preprocessor flag.

The WINDOWS_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to Windows operating systems.
The DARWIN_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to macOS operating systems.
The LINUX_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to Linux operating systems.
The WSL_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to Microsoft Subsystem for Linux.
Note: The WSL_ENABLED preprocessor macro is mostly useful for defining code sections that are written for compatibility with Microsoft Subsystem for Linux Version 1 (WSL1). For example, WSL1 is known to have issues with passing internal procedures to external routines since WSL has a non-executable stack. Although, this issue appears to have been resolved in WSL2, as a developer, you should be mindful of all users on all platforms.

Several issues also need special care when developing ParaMonte on Windows OS:

On Windows systems, file-locking by a processor often leads to problems and program crashes in parallel simulations. Since, currently only Intel compilers are supported and tested with the ParaMonte library, a quick remedy is activate the shared file manipulation via the Intel compiler SHARED argument that can be passed to open() statements. For example,

open( unit      = self%SampleFile%unit            &
    , file      = self%SampleFile%Path%original   &
    , status    = self%SampleFile%status          &
    , iostat    = self%SampleFile%Err%stat        &
#if INTEL_COMPILER_ENABLED && WINDOWS_ENABLED
    , SHARED                                      &
#endif
    , position  = self%SampleFile%Position%value  )

To properly build shared (DLL) libraries on Windows OS, every procedure name must be correctly extracted via Intel compiler DLLEXPORT directive. Currently, all names are manually extracted immediately below each procedure’s interface. This is normally not an issue, as long the symbols are correctly exported when the procedure is being developed. However, this can quickly become a challenge if the developer uses a wrong symbol name for the procedure or completely forgets to export the symbol.
Tip: A potential solution to this problem is to enable builds with CMAKE on Windows and request CMAKE to automatically export all symbols. Indeed, extending the current cmake build system of the ParaMonte library to Windows would be a great contribution to the package, if you can accomplish it. Before doing so, check the issues page of the project to make sure you would not duplicate someone else’s work.

Compiler suite preprocessor directives

Generally, one should avoid the use of code that bind the library to a particular compiler suite. Sometimes, however, this is unavoidable. Currently, the library supports builds with the Intel and GNU compiler suites.

The INTEL_COMPILER_ENABLED preprocessor macro must be defined with value 1 for any code section that requires the Intel compiler to recognize the syntax. This preprocessor flag is automatically defined by the ParaMonte build scripts and passed to the Intel preprocessor when the Intel compiler is used to build the library.
The GNU_COMPILER_ENABLED preprocessor macro must be defined with value 1 for any code section that requires the GNU compiler to recognize the syntax. This preprocessor flag is automatically defined by the GNU preprocessor on all platforms.

Parallelism preprocessor directives

There are currently two preprocessor directives that determine the type of parallelism to which the code section sandwiched by the preprocessor macro belongs.

The MPI_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to MPI-parallelized versions of the Fortran routines.
The CAF_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to CAF-parallelized versions of the Fortran routines (Coarray parallelism).
The OMP_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to OMP-parallelized versions of the Fortran routines (OpenMP parallelism). As of Dec 2020, there is no section of the Fortran routines uses OpenMP parallelism.

For example, if a section is MPI-parallelized, it must be fenced with the following preprocessor directive,

#if MPI_ENABLED
call mpi_initialized( isInitialized, ierrMPI )
#endif

or, if a section must be executed in either MPI or Coarray parallelisms, it must be fenced via the following preprocessor directive,

#if CAF_ENABLED || MPI_ENABLED
integer :: imageID
#endif

Library type preprocessor directives

Any code section that must be activated for a particular library build type (static vs. shared) must be fenced with appropriate preprocessing macros.

The DLL_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to shared library builds. Although DLL is reminiscent of Windows shared library files, the ParaMonte build scripts define this preprocessor macro for all shared library builds irrespective of the platform (Windows, Linux, macOS, …).
Important: Use this preprocessor macro in combination with IFORT_ENABLED to enable DLL symbol export by the Intel ifort compiler on macOS and Windows operating systems.

Library build preprocessor directives

Occasionally, it is necessary to define sections of code that should run only in a particular library build. Currently, the ParaMonte library build scripts support three different library builds (debug, testing, and release). Any code section that does not belong to the release (production) build must be appropriately fenced with appropriate preprocessor flags.

The DEBUG_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to the debug library builds. Use this macro to fence sections of code that help with the debugging of the code.
The TESTING_ENABLED preprocessor macro must be defined with value 1 for any code section that exclusively belongs to the testing library builds. Use this macro to fence sections of code that must be activated during the testing of the library.
Tip: The TESTING_ENABLED preprocessor macro often appears together with DEBUG_ENABLED and CODECOV_ENABLED. The testing mode deactivates all compiler optimizations, but also prevents the addition of the debugging symbols to the object files. This leads to faster compilation of the source files which is often desired for testing purposes.

Library interface preprocessor directives

To make the ParaMonte an inter-operable cross-language library, it is necessary to tailor sections of code for the particular languages of interest. The Fortran connection to all languages other than Fortran is provided via the FCI_ENABLED preprocessor macro in the ParaMonte library.

The CFI_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible from any programming language other than Fortran, most importantly, the C programming language.

In addition, some code sections might have to be exclusively executed when the library is invoked from a particular programming language. In such cases, the relevant code section must be properly fenced with the appropriate language-specific preprocessor macros.

The C_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the C programming language.
The CPP_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the C++ programming language.
The FORTRAN_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the Fortran programming language.
The JULIA_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the Julia programming language.
The MATLAB_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the MATLAB programming language.
The MATHEMATICA_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the Mathematica programming language.
The PYTHON_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the Python programming language.
The R_ENABLED preprocessor macro must be defined with value 1 for any section of the code that is meant to be accessible only when the library is built for the R programming language.

Coding style conventions

The following coding style are enforced within the Fortran files. If you do not follows these rules in your contribution, please provide an explanation and justification of the alternative approach that you have taken in your contribution.

Strict naming conventions are enforced within the entire library, including the Fortran routines.
Strict semantic compliance with the latest Fortran standard (2018).
Strict source-code compliance with the latest Fortran standard. However, there are important vital exceptions to this rule,
- It often preferable and sometimes essential to continue the source code lines beyond the current maximum line length limit specified by the standard, which is 132 characters. This maximum line length limit restriction is planned to be lifted in the next release of Fortran 202X standard. Nevertheless, a line length of approximately 200 characters is about where you should seriously think of breaking and continuing the code on multiple lines, if needed.
- Strict parallelism compliance with the Coarray Fortran parallelism is not required (via coarrays). In other words, you can implement parallelism that is provided by external libraries, the most prominent example of which is the Message Passing Interface (MPI) Library. However, in doing so, you will have properly fence the MPI section with the proper preprocessor macro as explained in the previous section.
- The minimal and efficient use of preprocessor macros and directives is allowed and encouraged to avoid code duplication and improve productivity.
All Coarray variables or objects (that are communicated between processes) must begin with the co_ prefix. Example: co_logFunc.
Why? This helps easily identify object that are responsible for inter-process communications in the Coarray implementation of the library.
All module variables or module objects must begin with mv_, unless there is a good justification for not doing so. Example: mv_state, comv_logFuncState.
Why? module variables are global, at least within the scope of the module, and at most, everywhere in the program. Global variables are extremely susceptible of creating extremely difficult hard-to-identify bugs in the code and are hard to follow. Adding the prefix mv_ to all global variables, resolves at least part of the problem by explicitly declaring the variable as global to the developer via its name.
All module constant variables that are supposed to be defined only once throughout the entire runtime must be prefixed with mc_ instead of mv_. Example: mc_imageCount = 3.
Why? Adding the prefix mc_ helps other developers to realize that the object is not supposed to change value at runtime beyond the first initialization. Note that this type of variable is different from a compile-time constant, since its value must be defined at runtime, which could be different from one run to another. However, once set, the value is expected to remain constant throughout the current simulation.
All constants (parameters) are upper-case separated by underscore. Example: FILE_EXT = ".txt".
All module names must begin with an uppercase letter and more importantly, must end with the prefix pm_ standing for the ParaMonte library. Example: pm_sampling.
Why? This helps create multiple different entity names from a single base name. For example,
```
use pm_str, only: str_type
type(str_type) :: str
```
Each module must be stored in a separate source file and the name of the source file must be the same as the module name, unless there is a good justification to do it otherwise. This is a strict rule to maintain the integrity of the library source file and the build scripts.
Modules should almost always contain only generic interfaces to functionalities.
If a module has submodules, the name of the submodule should follow the module name after @ in the actual file name containing the submodule. For example the following submodule,
```
submodule (pm_display) routines
...
end submodule routines
```
will have to be stored in a source file with the name pm_display@routines.f90.
Each module should generally associate with only one submodule named routines which contains the implementation of the generic interfaces in the parent module.
All type and class names must begin with an uppercase letter and more importantly, must end with the prefix _type. Example: sampler_type.
Why? This helps create multiple different entity names from a single base name. For example,
```
use pm_str, only: str_type
type(str_type) :: str
```

Comments

In general, the code itself should be expressive enough to obviate any need for comments. However, when needed, comments are encouraged and should be kept informative and concise.
Use the Doxygen convention and style for code annotations and documentation. See the ParaMonte Fortran source code for samples of Doxygen documentations of the Fortran routines.
If there is any work or comment to be left for the future developer(s) including yourself, use the syntax of Doxygen \todo (or less preferably @todo) to define future tasks and \author (or less preferably @author) to introduce yourself to the future developer within the code comments. For example,
```
!>  \todo
!>  \pvhigh
!>  There is room for performance improvements here by pre-computing the following array in procedure xxx.
```
The ParaMonte Doxygen documentation contains five predefined commands that indicate the importance of the task to do.
- \pvhigh : Priority Very High
- \phigh : Priority High
- \pmed : Priority Medium
- \plow : Priority Low
- \pvlow : Priority Very Low

Final steps

Once you have implemented your contributions,

Do not forget to test your contributions by adding new Fortran tests to the unit-testing framework of the library.
Also, generate code coverage report to ensure your contributions do not lower the overall code coverage of the Fortran routines.
Follow the generic contribution guidelines to submit and merge your contributions with the main branch of the library on GitHub.

If you have any questions about the topics discussed on this page, feel free to ask in the comment section below, or raise an issue on the GitHub page of the library, or reach out to the ParaMonte library authors.