The PM Programming Language is designed to facilitate the creation of numerical models on parallel systems. It combines concepts of parallelisation and vectorisation into a unified model using a 'strand' as a basic unit of parallelisation. A strand is a very lightweight entity that does not require its own stack and can map to a single execution of a loop body. Strands are used in both data and task parallelism and can communicate with other strands, including those running on other nodes. PM uses a modified form of the Partitioned Global Address Space approach which builds synchronisation into basic operations and excludes race conditions.
The PM compiler currently cross-compiles to Fortran+MPI. More target language/library combinations are planned, including Fortran-MPI-OpenMP and accelerator support (initially via either OpenMP or OpenACC)
The PM interpreter uses a parallel virtual machine and is designed for debugging PM code.
The language specification (available in /doc) while incomplete in places, will give a good view of the features of the version 0.4 of the language.
This is a pre-release version of the PM language. While syntax and semantics are now well developed, there may be further changes to both in response to feedback.
The implementation is not yet fully free of bugs or unimplemented features. If you encounter a problem, then please check the latest code on GitHub. If this does not work, then please raise an issue.
This is an open source project and outside contributions are entirely welcome.
At this stage the most effective contributions are in the are of testing and feedback on the language design and implementation.
A refactoring of compiler/interpreter source code is underway for version 0.5. One goal of this will be to make the code more accessible to outside contributors. In the meantime, bug-fix contributions to the source code by the brave are certainly welcome.
This is an initial release of version 0.4 of the language. At the moment the code is designed to be compiled and installed on a Linux system with MPI. It should be possible to compile on other systems with MPI and Fortan.
To compile PM:
Make sure you have MPI installed and loaded and can run mpifort and mpirun
The current setup is developed on Fedora with gfortran and openmp. You can edit pm/Makefile and pmc/Makefile to change compilers.
The code should be fairly portable. Portability issues may be addressed by editing config/sysdep.f90 (read the comments)
To compile the interpreter:
cd pm
make clean
make
You can then run the interpreter (./pm --help for options)
mpirun the interpreter for distributed execution.
To compile PM-to-Fortran compiler:
cd pmc
make clean
make
To compile a PM program to Fortran use ./pmc (./pmc --help for options)
The pmc compiler does not use MPI.
The compiled code will appear as PMOUT.F90 and will need to be compiled using mpifort.
You can then mpirun the resulting executable.
Both compiler and interpreter expect PM source files to have a .pmm extension.