Skip to content

Template for creating standalone FLAMEGPU2 models using the python3 interface

License

Notifications You must be signed in to change notification settings

FLAMEGPU/FLAMEGPU2-model-template-python

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 

Repository files navigation

FLAME GPU 2 Template for Python3

This repository can be used as a template for creating your own FLAME GPU 2 simulations or ensembles using the Python3 interface, using NVRTC, or the experimental native python approach.

For details on how to develop a model using FLAME GPU 2, refer to the userguide & API documentation.

C++ (CUDA) Interface

FLAME GPU 2 also provides a python-based interface for writing models. If you wish to use this instead of the python2 interface, see FLAMEGPU/FLAMEGPU2-model-template-cpp.

Dependencies

Getting pyflamegpu

pyflamegpu is currently available as pre-built python binary wheels, or can be built from source.

Pre-compiled pyflamegpu

Pre-built python wheels are available for Windows and Linux, for a range of Python versions on x86_64 systems. It is not currently available through any python package repositories.

To install a pre-built version of pyflamegpu:

  1. Download the appropriate .whl for the Latest Release
  2. Optionally create and activate a python venv or Conda environment
  3. Install the wheel locally via pip. See the release notes for details.

Building pyflamegpu

If the available python wheels are not appropriate for your system, or you wish to build with different CMake configuration options (i.e. FLAMEGPU_SEATBELTS=OFF for improved performance with reduced safety checks) you can build your own copy of pyflamegpu.

  1. Clone the main FLAMEGPU/FLAMEGPU2 git repository or download an archived release.
  2. Create a build directory and navigate to it.
  3. Configure CMake with FLAMEGPU_BUILD_PYTHON set to ON.
    • See the main FLAMEGPU/FLAMEGPU2 repository for further information on CMake configuration options
  4. Build the pyflamegpu target
  5. Optionally create and activate a python venv or Conda environment
  6. Install the wheel (from build/lib/<config>/python/dist/) locally via pip

Usage

Once pyflamegpu is installed into your local python installation or activated virtual environment, you can invoke the example model via:

python3 model.py <arguments>

Use -h/--help to see what command line arguments are available for the Simulation or Ensemble within the model

python3 model.py --help

Documentation and Support

For general information on FLAME GPU, Usage of FLAME GPU >= 2 and support see:

About

Template for creating standalone FLAMEGPU2 models using the python3 interface

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages