Different compilers in Pyccel

Different compilers in Pyccel#

Compilers supported by Pyccel#

Pyccel provides default compiler settings for 4 different compiler families:

  • GNU : gcc / gfortran

  • Intel : icx / ifx

  • NVIDIA : nvc / nvfort

  • LLVM: clang / flang

Warning : The NVIDIA compiler is not currently tested regularly

Specifying a compiler#

The default compiler family is GNU. To use a different compiler, the compiler family should be passed to either pyccel or epyccel. E.g.

pyccel example.py --compiler-family intel

or

epyccel(my_func, compiler_family='intel')

It is also possible to change the default compiler family by setting the environment variable PYCCEL_DEFAULT_COMPILER. E.g.

export PYCCEL_DEFAULT_COMPILER='intel'
pyccel example.py

The --compiler-family flag overrides the default compiler: if this is provided, the environment variable is ignored.

Using an unsupported compiler or custom compiler flags#

Pyccel’s compiler settings are described internally by a dictionary. This makes it easy for a user to define their own compiler settings. This is done via a JSON file. The location of the JSON file is passed to the --compiler-config flag of the pyccel command line interface, or to the compiler_config argument of the epyccel function. The JSON file must define the following parameters for each of the desired languages (we advise always including C in order to compile the wrapper):

  • exec : The name of the executable

  • mpi_exec : The name of the MPI executable

  • module_output_flag : This flag is only required when the language is Fortran. It specifies the flag which indicates where .mod files should be saved (e.g. ‘-J’ for gfortran)

  • debug_flags : A list of flags used when compiling in debug mode [optional]

  • release_flags : A list of flags used when compiling in release mode [optional]

  • general_flags : A list of flags used when compiling in any mode [optional]

  • standard_flags : A list of flags used to impose the expected language standard [optional]

  • libs : A list of libraries necessary for compiling [optional]

  • libdir : A list of library directories necessary for compiling [optional]

  • include : A list of include directories necessary for compiling [optional]

In addition, for each accelerator (mpi/openmp/openacc/python) that you will use the JSON file must define the following:

  • flags : A list of flags used to impose the expected language standard [optional]

  • libs : A list of libraries necessary for compiling [optional]

  • libdir : A list of library directories necessary for compiling [optional]

  • include : A list of include directories necessary for compiling [optional]

Python is considered to be an accelerator and must additionally specify shared_suffix.

The default compilers can provide examples compatible with your system once Pyccel has been executed at least. To export the JSON file describing your setup, use the --export-compiler-config flag and provide a target file name. E.g.

pyccel --compiler-family intel --export-compiler-config intel.json

once this file has been modified it can then be used with:

pyccel --compiler-config intel.json <file_to_translate>

Instead of using the --compiler-config flag, the environment variable PYCCEL_DEFAULT_COMPILER can be used to specify the path to the JSON file. This is especially useful in large projects where the pyccel command (or the epyccel Python function) is used many times with the same flags. E.g.

export PYCCEL_DEFAULT_COMPILER='intel.json'
pyccel mod1.py
pyccel mod2.py
pyccel subdir/mod3.py
...

Passing one of the flags --compiler-family or --compiler-config still allows the user to retrieve the normal behaviour of Pyccel.

Utilising Pyccel within the Anaconda environment#

While Anaconda is a popular way to install Python as it simplifies package management, it can introduce challenges when working with compilers.

Upon installation Anaconda modifies your shell’s environment variables.

This can lead to packages installed in conda/anaconda usurping the packages manually installed by the user.

This in turn easily leads to nasty surprises (e.g. using libraries compiled for serial execution instead of optimised parallel implementations).

To avoid these problems, Pyccel ignores Conda paths when searching for compilers in the system’s PATH.

By using the expected compiler, the expected libraries are correctly linked.

Pyccel offers the --conda-warnings flag which takes one of the following options : off, basic, verbose. This controls the visibility of Conda-related warnings. The default value is basic which indicates when folders are ignored. The verbose option additionally outputs a list of all ignored folders. If these folders should not be ignored then this can be avoided by using a user-defined compiler file as described above. The off option suppresses all Conda-related warnings.

Contents