7.2. OpenCL device selection

The pyopencl package supports selecting the device on which to run OpenCL code. There are multiple ways in which this selection can happen:

  • Interactively. When running mirgecom on the command line (i.e., not in a batch script), it will present a (potentially multiple-choice) dialog with the available OpenCL drivers (such as pocl) and the device (such as CPU or GPU). After a device has been selected, pyopencl will print the chosen device in the form of a value for the PYOPENCL_TEST environment variable that can be used for subsequent executions.

  • Through the PYOPENCL_TEST or PYOPENCL_CTX environment variables. Pyopencl will use the value of the PYOPENCL_TEST or PYOPENCL_CTX environment variables when set for device selection. The value of either variable can either be list indices (such as 0:1 for first driver, second device), or named abbreviations (such as port:tesla for pocl (=Portable OpenCL) and the first Nvidia Tesla GPU).

  • Default device. When not using one of the options above, pyopencl will run on the first available device (which might not be the device you want) by default.

Note

The device selection functionality described here is provided by the pyopencl.create_some_context(), pyopencl.tools.pytest_generate_tests_for_pyopencl(), and arraycontext.pytest_generate_tests_for_array_contexts() functions used in the default simulation drivers and tests. It is also possible to write your own device selection code with pyopencl.get_platforms(), pyopencl.Platform.get_devices(), and pyopencl.Context.

Note

Each MPI rank (=Python process) can only use one device during the execution, even if there are multiple devices available. Although it is possible to run multiple MPI ranks on the same device, we do not recommend doing this, as it will lead to contention.

Note

You can also use the clinfo command (automatically installed when installing emirge) to list all available OpenCL devices and their parameters.