FLAME GPU 2

FLAME GPU is a GPU accelerated agent-based simulation library for domain independent complex systems simulations. Version 2 is a complete re-write of the existing library offering greater flexibility, an improved interface for agent scripting and better research software engineering, with CUDA/C++ and Python interfaces.

FLAME GPU provides a mapping between a formal agent specifications with C++ based scripting and optimised CUDA code. This includes a number of key Agent-based Modelling (ABM) building blocks such as multiple agent types, agent communication and birth and death allocation.

• Agent-based (AB) modellers are able to focus on specifying agent behaviour and run simulations without explicit understanding of CUDA programming or GPU optimisation strategies.
• Simulation performance is significantly increased in comparison with CPU alternatives. This allows simulation of far larger model sizes with high performance at a fraction of the cost of grid based alternatives.
• Massive agent populations can be visualised in real time as agent data is already located on the GPU hardware.

Project Status

FLAME GPU 2 is currently in an pre-release (alpha) state, and although we hope there will not be significant changes to the API prior to a stable release there may be breaking changes as we fix issues, adjust the API and improve performance.

If you encounter issues while using FLAME GPU, please provide bug reports, feedback or ask questions via GitHub Issues and Discussions.

Documentation and Support

• Quickstart Guide
• Documentation and User Guide
• GitHub Discussions
• GitHub Issues
• Website

Installation

Pre-compiled python wheels are available for installations from Releases. Please see the latest release for more information on the available wheels and installation instructions.

C++/CUDA installation is not currently available. Please refer to the section on Building FLAME GPU.

Creating your own FLAME GPU Model

Template repositories are provided as a simple starting point for your own FLAME GPU models, with separate template repositories for the CUDA C++ and Python interfaces. See the template repositories for further information on their use.

• CUDA C++: FLAME GPU 2 example template project
• Python3: FLAME GPU 2 python example template project

Building FLAME GPU

FLAME GPU 2 uses CMake, as a cross-platform process, for configuring and generating build directives, e.g. Makefile or .vcxproj. This is used to build the FLAMEGPU2 library, examples, tests and documentation.

Requirements

Building FLAME GPU has the following requirements. There are also optional dependencies which are required for some components, such as Documentation or Python bindings.

• CMake >= 3.18
• CUDA >= 11.0 and a Compute Capability >= 3.5 NVIDIA GPU.
  • CUDA >= 10.0 currently works, but support will be dropped in a future release.
• C++17 capable C++ compiler (host), compatible with the installed CUDA version
  • Microsoft Visual Studio 2019 (Windows)
    • Note: Visual Studio must be installed before the CUDA toolkit is installed. See the CUDA installation guide for Windows for more information.
  • make and GCC >= 7 (Linux)
  • Older C++ compilers which support C++14 may currently work, but support will be dropped in a future release.
• git

Optionally:

• cpplint for linting code
• Doxygen to build the documentation
• Python >= 3.6 for python integration
• swig >= 4.0.2 for python integration
  • Swig 4.x will be automatically downloaded by CMake if not provided (if possible).
• FLAMEGPU2-visualiser dependencies
  • SDL
  • GLM (consistent C++/GLSL vector maths functionality)
  • GLEW (GL extension loader)
  • FreeType (font loading)
  • DevIL (image loading)
  • Fontconfig (Linux only, font detection)

Building with CMake

Building via CMake is a three step process, with slight differences depending on your platform.

1. Create a build directory for an out-of tree build
2. Configure CMake into the build directory
   • Using the CMake GUI or CLI tools
   • Specifying build options such as the CUDA Compute Capabilities to target, the inclusion of Visualisation or Python components, or performance impacting features such as SEATBELTS. See CMake Configuration Options for details of the available configuration options
3. Build compilation targets using the configured build system
   • See Available Targets for a list of available targets.

Linux

To build under Linux using the command line, you can perform the following steps.

For example, to configure CMake for Release builds, for consumer Pascal GPUs (Compute Capability 61), with python bindings enabled, producing the static library and boids_bruteforce example binary.

# Create the build directory and change into it
mkdir -p build && cd build

# Configure CMake from the command line passing configure-time options. 
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCH=61 -DBUILD_SWIG_PYTHON=ON

# Build the required targets. In this case all targets
cmake --build . --target flamegpu boids_bruteforce -j 8

# Alternatively make can be invoked directly
make flamegpu boids_bruteforce -j8

Windows

Under Windows, you must instruct CMake on which Visual Studio and architecture to build for, using the CMake -A and -G options. This can be done through the GUI or the CLI.

I.e. to configure CMake for consumer Pascal GPUs (Compute Capability 61), with python bindings enabled, and build the producing the static library and boids_bruteforce example binary in the Release configuration:

REM Create the build directory 
mkdir build
cd build

REM Configure CMake from the command line, specifying the -A and -G options. Alternatively use the GUI
cmake .. -A x64 -G "Visual Studio 16 2019" -DCUDA_ARCH=61 -DBUILD_SWIG_PYTHON=ON

REM You can then open Visual Studio manually from the .sln file, or via:
cmake --open . 
REM Alternatively, build from the command line specifying the build configuration
cmake --build . --config Release --target flamegpu boids_bruteforce --verbose

Configuring and Building a single example

It is also possible to configure and build individual examples as standalone CMake projects.

I.e. to configure and build game_of_life example in release mode from the command line, using linux as an example:

cd examples/game_of_life
mkdir -p build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCH=61
cmake --build . --target all

CMake Configuration Options

For a list of available CMake configuration options, run the following from the build directory:

cmake -LH ..

Available Targets

For a full list of available targets, run the following after configuring CMake:

cmake --build . --target help

Usage

Once compiled individual models can be executed from the command line, with a range of default command line arguments depending on whether the model implements a single Simulation, or an Ensemble of simulations.

To see the available command line arguments use the -h or --help options, for either C++ or python models.

I.e. for a Release build of the game_of_life model, run:

./bin/Release/game_of_life --help

Visual Studio

If wishing to run examples within Visual Studio it is necessary to right click the desired example in the Solution Explorer and select Debug > Start New Instance. Alternatively, if Set as StartUp Project is selected, the main debugging menus can be used to initiate execution. To configure command line argument for execution within Visual Studio, right click the desired example in the Solution Explorer and select Properties, in this dialog select Debugging in the left hand menu to display the entry field for command arguments. Note, it may be necessary to change the configuration as the properties dialog may be targeting a different configuration to the current build configuration.

Environment Variables

Several environmental variables are used or required by FLAME GPU 2.

• CUDA_PATH - Required when using RunTime Compilation (RTC), pointing to the root of the CUDA Toolkit where NVRTC resides.
  • i.e. /usr/local/cuda-11.0/ or C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.0.
  • Alternatively CUDA_HOME may be used if CUDA_PATH was not set.
• FLAMEGPU_INC_DIR - When RTC compilation is required, if the location of the include directory cannot be found it must be specified using the FLAMEGPU_INC_DIR environment variable.
• FLAMEGPU_TMP_DIR - FLAME GPU may cache some files to a temporary directory on the system, using the temporary directory returned by std::filesystem::temp_directory_path. The location can optionally be overridden using the FLAMEGPU_TMP_DIR environment variable.
• FLAMEGPU_RTC_INCLUDE_DIRS - A list of include directories that should be provided to the RTC compiler, these should be separated using ; (Windows) or : (Linux). If this variable is not found, the working directory will be used as a default.

Running the Test Suite(s)

To run the CUDA/C++ test suite:

1. Configure CMake with BUILD_TESTS=ON

2. Build the tests target

3. Run the test suite executable for the selected configuration i.e.

   ./bin/Release/tests

To run the python test suite:

1. Configure CMake with BUILD_SWIG_PYTHON=ON

2. Build the pyflamegpu target

3. Activate the generated python venv for the selected configuration, which has pyflamegpu and pytest installed

   If using Bash (linux, bash for windows)

   source lib/Release/lib/Release/python/venv/bin/activate

   If using cmd:

   call lib\Release\python\venv\Scripts\activate.bat

   Or if using Powershell:

   . lib\Release\python\venv\activate.ps1

4. Run pytest on the tests/swig/python directory. This may take some time.

   python3 -m pytest ../tests/swig/python

Contributing

Feel free to submit Pull Requests, create Issues or open Discussions.

See CONTRIBUTING.md for more detailed information on how to contribute to FLAME GPU.

Authors and Acknowledgment

See Contributors for a list of contributors towards this project.

If you use this software in your work, please cite DOI 10.5281/zenodo.5428984. Release specific DOI are also provided via Zenodo.

Alternatively, CITATION.cff provides citation metadata, which can also be accessed from GitHub.

License

FLAME GPU is distributed under the MIT Licence.

Known issues

There are currently several known issues which will be fixed in future releases (where possible). For a full list of known issues pleases see the Issue Tracker.

• Performance regressions in CUDA 11.3+, due to changes in compiler register usage (#560).
• Segfault when using flamegpu::DependencyGraph via the default constructor (#555). This will require an API break to resolve.
• Warnings and a loss of performance due to hash collisions in device code (#356)
• Multiple known areas where performance can be improved (e.g. #449, #402)
• Windows/MSVC builds using CUDA < 11.0 may encounter intermittent compiler failures. Please use CUDA 11.0+.
  • This will be resolved by dropping CUDA 10 support in a future release.
• Windows/MSVC builds using CUDA 11.0 may encounter errors when performing incremental builds if the static library has been recompiled. If this presents itself, re-save any .cu file in your executable producing project and re-trigger the build.
• Debug builds under linux with CUDA 11.0 may encounter cuda errors during validateIDCollisions. Consider using an alternate CUDA version if this is required (#569).
• CUDA 11.0 with GCC 9 may encounter a segmentation fault during compilation of the test suite. Consider using GCC 8 with CUDA 11.0.
• CMake 3.16 has known issues on some platforms. CMake versions less than 3.18 are deprecated and support will be removed in a future release.