The objective of this hackathon is to use Devito's JIT backdoor to transform the code that the Devito compiler generates for two different propagators. In particular:
- optimise the code automatically generated for
- OpenMP 5 offloading, or
- OpenACC;
- synthesize a new GPU implementation (e.g., CUDA, oneAPI) starting from the sequential version.
Code variants will be benchmarked, and the best performing strategies will be implemented in the Devito compiler. In this way, HPC/GPU developers can contribute to Devito development without having to learn how to develop the compiler itself.
Not able to make it to the OGHPC hackathon in person? No worries - we are a gloablly distributed dev team so everything is already set up to work online :-) At the top of our website, https://www.devitoproject.org/, you will see a link to our Slack workspace. Feel free to browse channels and ask questions - really the key channels for this hackathon are #oghpc and #gpu-dev.
You need to click on this link: https://labs.azure.com/register/jltt9mv2
You will be redirected to your Azure Labs VM. There you can set your password
for logging in to your VM. You have to wait for a minute or so. When your
password is updated, click on the screen emoji at your VM and select Connect via SSH
. Copy the ssh command to your terminal and start hacking!
Simply fork this repo. All forks will be checked periodically.
In the repository that you have just cloned, you will find two Python files.
One of these is run-preset.py
, which allows you to execute a set of Devito
benchmarks that we have pre-selected for this hackathon. Aspects like
discretization and duration of the benchmark are fixed, so you won't care about
them. The only argument that run-preset.py
accepts is the name of the
benchmark -- there are two alternatives:
python run-preset.py acoustic
: to generate and run an isotropic acoustic forward propagator.python run-preset.py tti
: to generate and run an anisotropic (TTI) acoustic forward propagator.
We suggest to start with the simplest one, acoustic
. The first time you run
the command, the code gets generated. Instructions will be provided on screen
explaining how to hack the generated code. Then, to test your hacking, just
re-run the same command. The execution time will be displayed at the end of
each run -- look for something along the lines of
Operator `Forward` run in 1.22 s
Where "Forward" is the name of the Operator. If interested in more performance
metrics, such as the GFlops/s and GPoints/s performance, just run with the
environment variable DEVITO_LOGGING=PERF
.
Note: we have configured the VM such that Devito, by default, generates and compiles code with OpenMP GPU-offloading pragmas. This is achieved through the following environment variables:
DEVITO_ARCH=clang
DEVITO_PLATFORM=nvidiaX
If one wants to:
- Use a different backend compiler, for example
pgcc
; then setDEVITO_ARCH=pgcc
and unsetDEVITO_PLATFORM
unset DEVITO_PLATFORM
export DEVITO_ARCH=pgcc
python run-preset.py acoustic
Devito will provide the location of the generated files. You may use vim/nano/emacs/... to modify the code, for example adding
- a new header files
#include "openacc.h"
, and #pragma acc kernels
to the parallel loops
then rerun:
python run-preset.py acoustic
to test your changes.
To submit your work at the end of the hackathon, just run python push-files.py
. And that's it!
Periodically we will:
- check all forks for updates
- checkout your updated code;
- run the benchmarks on a dedicated V100;
- update league table [will provide a link to the league table as soon as it is ready].
- We suggest that you always have a terminal running nvtop and/or htop. This can help not only to monitor the GPU usage of your VM but to quickly kill (SIGKILL) an application that is inefficient and save your time.
- Feel free to call the members of our team and ask for help if you consider any clarifications or more detailed explanation is needed.