GT4Py is a Python library for generating high performance implementations of stencil kernels from a high-level definition using regular Python functions. GT4Py is part of the GridTools framework, a set of libraries and utilities to develop performance portable applications in the area of weather and climate modeling.
NOTE: The gt4py.next
subpackage contains a new version of GT4Py which is not compatible with the current stable version defined in gt4py.cartesian
. The new version is highly experimental, it only works with unstructured meshes and it requires python >= 3.10
.
GT4Py is a Python library for expressing computational motifs as found in weather and climate applications. These computations are expressed in a domain specific language (GTScript) which is translated to high-performance implementations for CPUs and GPUs.
The DSL expresses computations on a 3-dimensional Cartesian grid. The horizontal axes (I
, J
) are always computed in parallel, while the vertical (K
) can be iterated in sequential, forward or backward, order. Cartesian offsets are expressed relative to a center index.
In addition, GT4Py provides functions to allocate arrays with memory layout suited for a particular backend.
The following backends are supported:
numpy
: Pure-Python backendgt:cpu_ifirst
: GridTools C++ CPU backend using I
-first data orderinggt:cpu_kfirst
: GridTools C++ CPU backend using K
-first data orderinggt:gpu
: GridTools backend for CUDAcuda
: CUDA backend minimally using utilities from GridToolsdace:cpu
: Dace code-generated CPU backenddace:gpu
: Dace code-generated GPU backendGT4Py can be installed as a regular Python package using pip
(or any other PEP-517 frontend). As usual, we strongly recommended to create a new virtual environment to work on this project.
The performance backends also require the Boost library, a dependency of GridTools C++, which needs to be installed by the user.
If GridTools or Boost are not found in the compiler's standard include path, or a custom version is desired, then a couple configuration environment variables will allow the compiler to use them:
GT_INCLUDE_PATH
: Path to the GridTools installation.BOOST_ROOT
: Path to a boost installation.Other commonly used environment variables are:
CUDA_ARCH
: Set the compute capability of the NVIDIA GPU if it is not detected automatically by cupy
.CXX
: Set the C++ compiler.GT_CACHE_DIR_NAME
: Name of the compiler's cache directory (defaults to .gt_cache
)GT_CACHE_ROOT
: Path to the compiler cache (defaults to ./
)More options and details are available in config.py
.
GT4Py uses Sphinx documentation. To build the documentation install the dependencies in requirements-dev.txt
pip install -r ./gt4py/requirements-dev.txt
and then build the docs with
cd gt4py/docs/user/cartesian
make html # run 'make help' for a list of targets
Follow the installation instructions below to initialize a development virtual environment containing an editable installation of the GT4Py package. Make sure you read the CONTRIBUTING.md and CODING_GUIDELINES.md documents before you start working on the project.
tox
If tox is already installed in your system (tox
is available in PyPI and many other package managers), the easiest way to create a virtual environment ready for development is:
# Clone the repository
git clone https://github.com/gridtools/gt4py.git
cd gt4py
# Create the development environment in any location (usually `.venv`)
# selecting one of the following templates:
# dev-py310 -> base environment
# dev-py310-atlas -> base environment + atlas4py bindings
tox devenv -e dev-py310 .venv
# Finally, activate the environment
source .venv/bin/activate
Alternatively, a development environment can be created from scratch installing the frozen dependencies packages :
# Clone the repository
git clone https://github.com/gridtools/gt4py.git
cd gt4py
# Create a (Python 3.10) virtual environment (usually at `.venv`)
python3.10 -m venv .venv
# Activate the virtual environment and update basic packages
source .venv/bin/activate
pip install --upgrade wheel setuptools pip
# Install the required development tools
pip install -r requirements-dev.txt
# Install GT4Py project in editable mode
pip install -e .
# Optionally, install atlas4py bindings directly from the repo
# pip install git+https://github.com/GridTools/atlas4py#egg=atlas4py
GT4Py is licensed under the terms of the BSD-3-Clause.