Getting Started
Installation
Using pip from PyPI:
python3 -m pip install pyaccl
Locally from source distribution:
python3 setup.py sdist && python3 -m pip install dist/*.tar.gz
Locally editable:
python3 -m pip install -e .
Downloading Notebooks and Overlays
PyACCL provides a few Jupyter notebooks and associated FPGA overlays for Alveo boards. After installing the PyACCL package, you can download these with:
pynq get-notebooks --from-package pyaccl all
Pynq will automatically download the right overlay for the Alveo
device(s) in your system. If you do not have an Alveo board, add
--ignore-overlays to the above command. The notebooks and overlays
will be downloaded to the folder where the command was executed, and a
Jupyter notebook server can be started from there.
There are several notebooks available to get you started with PyACCL:
Intros to (Py)ACCL primitives and collectives
Short guides to using compression and communicators
Quick overview of performance-related flags
Running Tests
PyACCL includes tests for single ACCL instances and systems of ACCL instances. The tests are designed to run against ACCL emulator/simulator sessions as well as the single-FPGA ACCL test overlay, which connects 3 ACCL instances on a single Alveo board.
To run the tests, add the ACCL emulator executable cclo_emu to your
path and start your emulator/simulator session:
pyaccl-emulate -n <NRANKS>
then run the following command from the pyaccl root folder:
mpirun -np <NRANKS> python3 -m pytest --with-mpi
If your system has an Alveo board and NRANKS is less or equal to 3,
the test fixture will try and download the appropriate overlay for it,
otherwise it will skip hardware testcases. Similarly, the test fixture
will attempt to identify a valid emulator/simulator session with the
appropriate number of ranks, then run testcases against it. If you omit
--with-mpi, only single-instance tests will run, i.e. tests for
copy() and combine() primitives.