WODEN is built for speed and only works with a GPU. Currently, you need either an NVIDIA GPU to use CUDA functionality, or something else that can use HIP (likely an AMD GPU). CUDA is tried and tested, whereas HIP is new in version 2.2 and not well tested. Furthermore, WODEN has only been tested to run on linux, specifically Ubuntu 16.04 up to 24.04. This does however include the Windows Subsystem for Linux 2 (WSL 2)., so you can technically run in on Windows kinda.
You have two options for installation:
More work, but tailored to your system: Manual Installation
Less work, but less flexibility/performance: Use a Docker image
Both options are described below, jump to whatever suits you.
WODEN has been tested to run on the following Australian super computers:
Garrawarla (Pawsey) CUDA
OzStar (Swinburne University) CUDA
Ngarrgu Tindebeek (Swinburne University) CUDA
Setonix (Pawsey) HIP
Manual Installation¶
For examples of building WODEN from source on superclusters, see:
WODEN/templates/install_woden_nt.shfor a CUDA build on Ngarrgu TindebeekWODEN/templates/install_woden_setonix.shfor a HIP build on Setonix
Dependencies¶
WODEN has a number of dependencies so it doesn’t reinvent the wheel. A brief list of them here is followed by detailed instructions on how I installed them in the following subsection.
CMake - https://cmake.org version >= 3.21
Either NVIDIA CUDA - https://developer.nvidia.com/cuda-downloads
or AMD ROCm - https://rocm.docs.amd.com/projects/install-on-linux/en/latest/
rust - https://www.rust-lang.org/tools/install (needed for
mwa_hyperbeam)mwa_hyperbeam - https://github.com/MWATelescope/mwa_hyperbeam
Python >= 3.8 (as well as a number of Python modules, see below)
How to install dependencies¶
These instructions are for Ubuntu 24.04, but can be used as a guide for otherlinux-like systems.
CMake - https://cmake.org version >= 3.21:
$ sudo snap install cmake
NVIDIA CUDA - https://developer.nvidia.com/cuda-downloads. Best used if you have an NVIDIA GPU. I typically download the runfile option, which you run as:
$ sudo sh cuda_11.2.2_460.32.03_linux.run ##your version will likely be different
but I do NOT install the drivers at this point, as I’ll already have drivers. Up to you and how your system works. Also, don’t ignore the step of adding something like
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-11.2/lib64to your~/.bashrc, or your system won’t findCUDA.AMD ROCm - https://rocm.docs.amd.com/projects/install-on-linux/en/latest/:
I don’t have an AMD GPU, so I’ve never done this. Fingers crossed the linked instructions work for you!
mwa_hyperbeam - https://github.com/MWATelescope/mwa_hyperbeam -
mwa_hyperbeamis the go-to package for calculating the MWA Fully Embedded Element (FEE) primary beam model. At the time of writing (23/03/2022), we’ll have to install and compile from source to get the CUDA code that we want to link to. We should be able to install release versions in the future. For now, you’ll first need to installrust, the language the library is written in. I followed the installation guide at https://www.rust-lang.org/tools/install, which for me on Ubuntu just means running:$ curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
Once that’s installed, I run the following commands for a CUDA installation (you can choose where to install it, I’m just putting where I happened to do it this time round):
$ cd /home/jline/software$ git clone https://github.com/MWATelescope/mwa_hyperbeam.git$ cd mwa_hyperbeam$ export HYPERDRIVE_CUDA_COMPUTE=60 ##your compute capability$ cargo build --locked --release --features=cuda,hdf5-static
Note
export HYPERDRIVE_CUDA_COMPUTE=60is not essential as the compiler should be smart enough, but you might get a speed boost but setting the correct architecture. This of course depends on your GPU; see ‘Machine specifics’ below on how to work out your architecture.If you have an AMD GPU, replace the last two lines with something like:
$ export HYPERBEAM_HIP_ARCH=gfx90a$ cargo build --locked --release --features=hip,hdf5-static
where again the value of
HYPERBEAM_HIP_ARCHdepends on what kind of GPU you have.That’s it! I’ll show you how to link to it later when we install
WODEN. If you don’t want to have to tellCMakewhere to look for the libraries, you’ll need to link/copylibmwa_hyperbeam.sosomewhere your compiler can see, as well asmwa_hyperbeam.h.python >= 3.8 - 3.8 should work, but I’d suggest going with 3.11 or 3.12. The requirements can be found in
WODEN/requirements.txt, which you can install via something like:$ pip3 install -r requirements_testing.txt
For completeness, those packages are:
sphinx_argparsebreatheastropy<=6.1.0numpy<=1.26.0pyerfapalpyimportlib_resourcessphinx-math-dollarmatplotlibpyuvdatapython-casacore
Phew! That’s it for now.
Compiling WODEN C/CUDA code¶
In an ideal world, if the installation of your dependencies went perfectly andyou have a newer NVIDIA GPU, you should be able to simply run:
$ git clone https://github.com/JLBLine/WODEN.git$ cd WODEN$ mkdir build && cd build$ cmake ..$ make -j 4
et voila, your code is compiled. Keep reading to see how to install WODEN so you can run it from anywhere.
Warning
Even if the code compiled, if your GPU has a compute capability < 5.1, newer versions of nvcc won’t compile code that will work. You’ll get error messages like “No kernel image available”. Check out how to fix that in ‘Machine specifics’ below.
Machine specifics¶
It’s almost a guarantee cmake won’t be able to find mwa_hyperbeam, so you’ll have to point it to where things are installed. You can use two keywords in the following way to achieve that:
$ cmake .. -DHBEAM_INC=/home/jline/software/mwa_hyperbeam/include \ -DHBEAM_LIB=/home/jline/software/mwa_hyperbeam/target/release/libmwa_hyperbeam.so
Obviously you’ll need to point to where you have installed things. If you have a library with my name in the path I’d be concerned, so edit it as appropriate.
All NVIDIA GPUs have a specific compute capability, which relates to their internal architecture. You can tell the compiler which architecture to compile for, which in theory should make compilation quicker, and ensure the code runs correctly on your GPU. You can find out the compute value here (https://developer.nvidia.com/cuda-gpus), and pass it to CMake via setting the CUDAARCHS environment variable (https://cmake.org/cmake/help/latest/envvar/CUDAARCHS.html) BEFORE you run the call to cmake:
$ export CUDAARCHS=60
(for a compute capability of 6.0, for example).
Warning
For newer CUDA versions, some compute capabilities are deprecated, so the compiler leaves them out by default. For example, using CUDA version 11.2, compute capabilities 3.5 to 5.0 are ignored. If you card has a compute capability of 5.0, you must include the flag -DCUDA_ARCH=5.0, otherwise the nvcc compiler will not create an executable capable of running on your device.
If you need to pass extra flags to your CUDA compiler, you can do so by adding something like the following (noting that all CMake flags start with -D):
-DCMAKE_CUDA_FLAGS="-Dsomeflag"
Compiling WODEN C/HIP code¶
If you have an AMD GPU, you can compile the HIP code instead of the CUDA code. This is a new feature in WODEN and not as well tested. You can compile the HIP code by setting the USE_HIP flag to ON when you run cmake (you’ll still need to link ):
$ cmake .. -DUSE_HIP=ON \ -DHBEAM_INC=/home/jline/software/mwa_hyperbeam/include \ -DHBEAM_LIB=/home/jline/software/mwa_hyperbeam/target/release/libmwa_hyperbeam.so
Machine specifics¶
Similarly to CUDA, you can set a HIP architecture. To find out which one you need, try:
$ offload-arch
which spat of gfx90a for me. You pass that onto cmake via the HIP_ARCH flag:
$ cmake .. -DUSE_HIP=ON -DHIP_ARCH=gfx90a \ -DHBEAM_INC=/home/jline/software/mwa_hyperbeam/include \ -DHBEAM_LIB=/home/jline/software/mwa_hyperbeam/target/release/libmwa_hyperbeam.so
Fair warning, I had to include the HIP_ARCH flag. The code would compile fine but not work at runtime, so a bit nasty.
Installing wodenpy¶
OK, we’ve compiled the C/GPU libraries; now to install the WODEN Python package and executables. You can do this by running:
$ cd WODEN$ pip3 install .
That’s it. You should be able to run run_woden.py --help on the command line.
Post compilation (optional)¶
If you want to use the MWA FEE primary beam model, you must have the stored spherical harmonic coefficients hdf5 file mwa_full_embedded_element_pattern.h5. You can then define this environment variable in your ~/.bash_rc:
export MWA_FEE_HDF5=/path/to/your/location/mwa_full_embedded_element_pattern.h5
so run_woden.py can find it. There is a command line option --hdf5_beam_path in run_woden.py which you can use instead of this environment variable if you want.
If you don’t have the spherical harmonic file you can obtain it via the command:
$ wget http://ws.mwatelescope.org/static/mwa_full_embedded_element_pattern.h5
To use the interpolated MWA FEE beam model, do similarly:
$ wget http://ws.mwatelescope.org/static/MWA_embedded_element_pattern_rev2_interp_167_197MHz.h5$ export MWA_FEE_HDF5_INTERP=/path/to/your/location/MWA_embedded_element_pattern_rev2_interp_167_197MHz.h5
Use a Docker image¶
Note
All the images listed here were created with the script WODEN/docker/make_docker_image.sh. If a particular image doesn’t work for you, you can edit the source to hopefully get it working.
For CUDA¶
Fair warning, this is a new option, and hasn’t been heavily tested. I have successfully run it on a number of clusters (via singularity). Which version you pull depends on your GPU. If you have an NVIDIA GPU, you need to work out what your compute capability is, and pull the appropriate image. Say you have an NVIDIA V100 card, you have a compute capacity of 7.0, so you’d pull the image like this:
$ docker pull jlbline/woden-2.3:cuda-70
I have made images for computes 60,61,70,75,80,86. If you need another compute, either run the Docker script to make a new docker, or just compile from source as instructed above. In theory, you can just run WODEN commands by doing something like this:
$ docker run -it --gpus all woden-2.3:cuda-70 \ --env XDG_CONFIG_HOME=/somewhere/astropy_storage \ --env XDG_CACHE_HOME=/somewhere/astropy_storage \ run_woden.py --help
where the --gpus all means the docker instance can see your GPUs. The environment variables point to somewhere to keep your astropy outputs, which is useful if you’re running somewhere you’re not admin (like on a cluster). There must be a better way to do this but I’m a docker noob.
For HIP¶
The only HIP image I’ve made is for the Setonix cluster, and is based on a Pawsey specific image https://quay.io/repository/pawsey/rocm-mpich-base?tab=tags&tag=latest. You can pull it like this:
$ docker pull jlbline/woden-2.3:setonix
It is highly unlikely it won’t work anywhere else.
Using singularity¶
For CUDA¶
If your system has singularity and not docker, you can convert the docker image to a singularity image via:
$ singularity build woden-2.3-70.sif docker://jlbline/woden-2.3:cuda-70
with an example of running the help looking something like:
$ singularity exec --nv --home=/astro/mwaeor/jline \ woden-2.3-70.sif run_woden.py --help
Similarly to the docker image, --nv means use the NVIDIA GPUs, and --home sets a specific location to treat as home if you’re not on a local machine.
For HIP¶
Again, the only HIP image I’ve made is for the Setonix cluster where you can do:
$ singularity build woden-2.3:setonix.sif docker://jlbline/woden-2.3:setonix
and run it like:
$ singularity exec --home=/scratch/mwaeor/jline \${MYSOFTWARE}/woden-2.3-setonix.sif run_woden.py --helpWarning
EVERYTHING on the internet will tell you to use the --rocm flag. This WILL NOT WORK with the Setonix based image, because of shenanigans. So leave it be.
