Installation

katsdpsigproc is available from the Python Package Index, so it can be installed just with:

pip install katsdpsigproc

The package itself is pure Python, but a number of the dependencies have compiled code and may require a compiler to be present if you have a platform that isn’t supported by the available binary wheels.

Alternatively, the latest development version can be installed from Github with:

pip install https://github.com/ska-sa/katsdpsigproc

This is sufficient for running the CPU code, but for acceleration using CUDA or OpenCL you will also need to install pycuda or pyopencl respectively.

Configuration

To check that your drivers have been successfully detected, run the following script:

#!/usr/bin/env python
import katsdpsigproc.accel

ctx = katsdpsigproc.accel.create_some_context(interactive=True)
print(f"Successfully created context on {ctx.device.name} ({ctx.device.platform_name})")

If more than one device is detected, it will list them and prompt you to select one. This can happen even if you only have a single GPU, if it is available both via CUDA and OpenCL. Note that for OpenCL the “platform name” is shown in parentheses; for NVIDIA GPUs this is confusingly “NVIDIA CUDA”. If you want to use the device via the CUDA API, it is the one with just “CUDA” in parentheses.

An environment variable can be used to specify a device so that you are not prompted every time:

KATSDPSIGPROC_DEVICE

Specify the device to use. The value must be a number, which matches the numbering offered by the menu.

Troubleshooting

CUDA on Linux

If the script above does not work or does not detect a CUDA device, the following steps will help isolate the problem.

  1. Check that pycuda is installed. If it is not, you won’t get an error, but CUDA devices will not be available.

  2. Run nvidia-smi. This checks that the NVIDIA drivers are installed and that the userspace component can talk to the kernel component. If not, you might need to reinstall the drivers.

  3. Run a compiled CUDA program. For example, the CUDA installation will usually offer to install samples. The deviceQuery sample is a good test since it’ll show the devices detected. If this fails then there is likely something wrong with your CUDA library installation.

  4. Run python -c 'import pycuda.autoinit'. If that fails (and the previous steps succeed) the error is most likely into your pycuda installation. A common problem is that CUDA is upgraded but pycuda is still linked against the old version. In this case you need to uninstall it, remove any cached binary wheel, and reinstall it.